In this episode of the M365.fm podcast, we dive deep into building and deploying production-grade AI agents with Microsoft Foundry together with Edgar McOchieng (MVP). The conversation explores how organizations can move beyond AI demos and prototypes into scalable, secure, and enterprise-ready agentic solutions.

Edgar shares practical insights into Microsoft Foundry, Azure AI services, orchestration patterns, governance, observability, and the challenges teams face when deploying AI agents in real-world environments. We discuss how developers can combine tools like Azure OpenAI, Semantic Kernel, MCP integrations, and multi-agent architectures to create intelligent systems that are reliable, maintainable, and aligned with enterprise requirements.

The episode also covers key topics such as memory management, tool integration, security boundaries, human-in-the-loop workflows, responsible AI practices, and monitoring strategies for production workloads. Edgar explains why architecture, governance, and operational excellence are just as important as model selection when building AI-powered business applications.

Whether you are an AI engineer, Microsoft 365 professional, architect, or developer exploring the future of agentic AI, this episode provides actionable guidance and real-world lessons for designing modern AI solutions with Microsoft Foundry and the Microsoft ecosystem.

Tune in to learn how enterprises can successfully build, deploy, scale, and govern AI agents that deliver real business value.

You may find that building ai agents for demos is easy, but moving them to real-world use demands more. Production grade AI agents require you to focus on reliability, scalability, and governance.

Reliability shifts from simple validation to robust, sustainable performance.
Scalability becomes a challenge, as few prototypes reach enterprise scale due to integration and security needs.
Governance must follow strict frameworks like Zero-Trust to meet compliance.

Microsoft Foundry helps you break down data silos, making your information both accessible and trustworthy. This unlocks strategic advantages, such as cost savings and new revenue streams. You can harness state-of-the-art language models to power innovation across industries.

Key Takeaways

Building production-grade AI agents requires a focus on reliability, scalability, and governance.
Microsoft Foundry breaks down data silos, making information accessible and trustworthy for better decision-making.
Utilize the model playground and no-code tools to experiment with AI models without needing deep coding skills.
Integrate Microsoft Foundry with existing Microsoft 365 apps to enhance productivity and streamline workflows.
Adopt a modular approach with the Microsoft agent framework to build flexible and scalable AI solutions.
Implement strong governance and security measures to protect data and ensure compliance with industry standards.
Monitor AI agents in real-time to maintain performance and quickly address any issues that arise.
Follow best practices for development and deployment to ensure your AI agents are reliable, secure, and efficient.

Microsoft Foundry Platform Overview

Core Capabilities

You can unlock a wide range of features with the microsoft foundry agent service. The platform gives you access to model experimentation, foundry hosted agents, and orchestration pipelines. These tools help you move from simple prototypes to robust, enterprise-ready solutions. The microsoft foundry agent service supports you with real-time monitoring, security, and business context awareness. You can manage data, track usage, and ensure compliance with ease.

Here is a quick look at the main features you get with the microsoft foundry agent service:

Feature	Description
Model Hub & Deployment	Pre-trained models, custom fine-tuning, A/B testing, automatic failover, and load balancing.
Data & Vector Management	Integrated vector databases, data preprocessing, automated lineage tracking, GDPR compliance.
Security & Governance	Role-based access control, content filtering, audit logging, private endpoints.
Monitoring & Observability	Real-time metrics, cost tracking, usage analytics, automated alerts.
Zero Infrastructure Management	Fully managed Azure services with automatic scaling and global availability.
Integrated Toolchain	End-to-end ML/AI workflow with built-in compliance and security.
Customization & Upgrades	Train, fine-tune, and upgrade models with minimal coding.
Orchestration	Streamlined development with open frameworks and serverless flexibility.
Business Context Awareness	Agents can securely connect to multiple data sources and automate business processes.

You can see how the microsoft foundry agent service stands out by offering a fully managed platform that covers every step of the AI lifecycle.

Model Playground and No-Code Tools

With the microsoft foundry agent service, you can experiment with different models in the model playground. This space lets you test, compare, and fine-tune models before deploying them as foundry hosted agents. You do not need deep coding skills to get started. The no-code tools help you build, configure, and launch foundry hosted agents quickly. You can drag and drop components, set up workflows, and monitor performance—all from a simple interface.

The model playground also supports A/B testing and automatic failover. This means you can ensure your foundry hosted agents deliver reliable results. You can upgrade or retrain your agents as your needs change, keeping your solutions up to date.

Integration with Microsoft and External Services

You can connect the microsoft foundry agent service to Microsoft 365 apps like SharePoint, OneDrive, and Outlook. The platform also links with AWS and Google Cloud, making it easy to break down data silos. Microsoft Purview helps you manage policies and data visibility from one place. This centralized approach ensures you keep your data secure and compliant, even when you use multiple cloud providers.

By using foundry hosted agents, you can automate business processes across different platforms. The microsoft foundry agent service lets you analyze data efficiently while protecting sensitive information. You can give your teams self-service access to the data they need, boosting productivity and supporting better decisions.

Tip: When you use foundry hosted agents, you can streamline your workflows and reduce manual tasks. This helps your organization move faster and stay competitive.

Architecture for Production Grade AI Agents

Modular Components

You need a flexible foundation when you build production grade ai agents. The microsoft agent framework gives you modular components that fit together like building blocks. You can choose the tools you need and connect them to your business systems. This open and composable structure lets you add custom database queries or internal APIs. You can deploy agents to edge devices or private clouds, which gives you more control and flexibility.

You can integrate custom tools and APIs with the microsoft agent framework.
You can deploy agents to edge devices or private clouds.
Cloud-edge collaboration helps you manage modern AI applications.
The microsoft agent framework supports complex multi-agent scenarios.
Built-in observability and compliance features help you monitor and manage agents.
Standardized protocols like MCP and A2A make integration with Microsoft services seamless.

You can use the microsoft agent framework to build production grade ai agents that adapt to your needs. You can scale up or down, add new features, and keep your agents running smoothly.

Hybrid AI Architecture

You can use hybrid AI architecture to make your production grade ai agents more powerful. The microsoft agent framework lets you switch between local and cloud-hosted models. You can choose the best model for each task. This approach helps you optimize costs and performance. You can ground your agents in enterprise data and connect them to business systems. Orchestration across multi-agent teams becomes easier with the microsoft agent framework.

Note: Hybrid AI architecture gives you the flexibility to use specialized models for different tasks. You can keep your agents reliable and efficient.

You can use the microsoft agent framework to build agents that work together. Centralized governance ensures you can innovate without losing control. You can manage permissions, monitor usage, and keep your data secure.

Scalability and Reliability

You need your production grade ai agents to scale and stay reliable. The microsoft agent framework uses modular design and event-driven autoscaling. You can align resource consumption with business demand. Specialized models help your agents focus on specific tasks. A robust knowledge layer protects against model hallucination and keeps your agents accurate.

Feature	Contribution to Scalability and Reliability
Modular Design	You can manage and scale components independently, which improves reliability.
Event-Driven Autoscaling	You can match resources to business demand, which keeps performance steady under load.
Specialized Models	You can assign agents to specific tasks, which boosts reliability and performance.
Robust Knowledge Layer	You can ensure agents use accurate, up-to-date information, which prevents errors.
Comprehensive Observability	You can track AI-specific KPIs, which helps you manage operational health and business value.

You can use the microsoft agent framework to monitor your agents in real time. You can track metrics, set automated alerts, and analyze usage. You can keep your production grade ai agents running at peak performance. You can upgrade or retrain agents as your needs change. You can rely on the microsoft agent framework to deliver enterprise-ready solutions.

Orchestrating and Managing AI Agents

Multi-Agent Workflows

You can orchestrate intelligent ai agents to handle complex business processes. Microsoft Foundry lets you design systems where each agent has a specialized role. This approach helps you avoid relying on a single agent for every task. You can build stateful workflows that coordinate agents across long-running processes. These workflows support durability and error recovery, so your operations stay resilient.

You can use multi-agent orchestration to assign tasks to agents with unique skills.
You can embed governance into your workflows from the start, ensuring compliance and trust.
You can assign a unique identity to each agent, which allows for controlled access and auditing.

When you use multi-agent workflows, you can streamline tasks like procure-to-pay. Specialized agents can reduce cycle time from weeks to hours. You can synchronize supply chains across organizations, enabling rapid coordination among logistics and customs agents. This minimizes delays and keeps your business moving.

Tip: Multi-agent workflows help you automate routine tasks and free up your team for more strategic work.

Agent Collaboration

You can foster collaboration among intelligent ai agents to boost productivity. Microsoft Foundry enables agents to work together, sharing information and coordinating actions. Agents can augment knowledge workers by managing scheduling and routine tasks. This allows your team to focus on complex decisions.

You can use agent collaboration to improve IT operations and incident response. Agents detect anomalies and diagnose issues quickly.
You can automate cross-team communication, making it easier to resolve problems and share insights.

Collaboration among agents helps you achieve end-to-end automation. You can create workflows where agents support each other, ensuring smooth handoffs and consistent results.

Monitoring and Observability

You need to monitor and observe your AI agents to ensure reliability. Microsoft Foundry provides deep visibility into agent behavior and decision-making. You can use dashboards to track operational metrics in real time. Built-in evaluators measure the quality, safety, and reliability of agent responses. Tracing tools capture the execution flow, helping you debug and analyze performance.

Capability	Description
Evaluation	Measures quality, safety, and reliability of AI responses with built-in and custom evaluators.
Monitoring	Ensures quality and performance in production with real-time dashboards tracking various operational metrics.
Tracing	Captures execution flow of AI applications for debugging and performance analysis.

You can use these tools to gain confidence in your AI systems. Observability helps you identify issues early and maintain operational health. You can audit agent actions and ensure compliance with enterprise standards.

Note: Monitoring and observability are essential for maintaining trust and transparency in your AI environment.

Governance, Security, and Compliance

Building AI agents for enterprise use requires you to focus on strong governance, security, and compliance. Microsoft Foundry gives you a set of tools and features that help you meet industry standards and protect your organization’s data.

Permissions and Access Control

You need to control who can access your AI agents and data. Microsoft Foundry uses advanced identity and access management to keep your environment secure. You can assign roles to users and agents, making sure each one only gets the permissions needed for their tasks. This approach follows the principle of least privilege, which is important for enterprise use.

Here is a table that shows how Microsoft Foundry manages permissions and access:

Feature	Description
RBAC Implementation	Assign built-in roles like Azure AI User or Project Manager to enforce least privilege.
Managed Identity	Use system-assigned identities for agents to access services without storing secrets.
Logging and Traceability	Track every action back to a user or agent identity for full accountability.

You can also use Microsoft Entra authentication and managed identities. This lets you avoid static API keys and improves security. The SDK supports access tokens and delegated consent, so you can review and manage permissions easily.

Auditability and Transparency

You must be able to track and understand what your AI agents do. Microsoft Foundry gives you deep observability and logging. Every agent gets a unique identity, so you can trace actions and decisions back to the source. This helps you meet compliance needs and build trust with stakeholders.

You can map agent risks and impacts to ensure actions are well-considered.
Human oversight lets you verify and approve important actions.
The system defines clear boundaries for agent operations, making it easier to understand and troubleshoot.
Centralized dashboards show you agent behavior and help you monitor compliance.

Microsoft Foundry uses integrated logging and telemetry tools like Azure Monitor and Application Insights. These tools help you keep a record of every action, which is essential for audits and regulatory checks.

Responsible AI Practices

You need to make sure your AI agents act responsibly and follow ethical guidelines. Microsoft Foundry supports responsible AI by giving you tools for data security, policy enforcement, and risk management.

Microsoft Purview integrates with Foundry to secure and govern AI interactions.
You can classify data, apply consistent policies, and audit agent activities.
Real-time threat protection helps you detect risks and vulnerabilities.
The platform uses a layered security model to protect sensitive information.

Microsoft Foundry’s compliance features help you meet regulations like GDPR and HIPAA. The platform’s strong security controls, such as tenant isolation and customer data usage policies, make it a reliable choice for regulated industries. You can compare these features to other platforms, but Foundry stands out for its deep integration with Azure security tools and its focus on enterprise use.

Note: Responsible AI practices are not just about technology. You must combine strong tools with human oversight to ensure safe and ethical outcomes.

Seamless Enterprise Integration

Connecting to SharePoint, OneDrive, Outlook

You can connect Microsoft Foundry directly to SharePoint, OneDrive, and Outlook. This integration helps you unlock the value of your organization’s data. You do not need to move files or emails to a new system. Foundry works with your existing Microsoft 365 environment. You can search, analyze, and automate tasks using the data you already have.

Foundry recognizes SharePoint access control lists automatically. This means you do not have to set up new permissions for your AI agents. Your data stays secure, and you keep control over who can see what. You can use Microsoft Copilot Studio to add advanced AI features to your workflows. This makes it easy to build smart solutions that fit your business needs.

You can also automate repetitive tasks with Power Automate. Foundry supports over 1,400 connectors, so you can link your agents to many different apps and services. This saves you time and reduces manual work.

Tip: When you connect Foundry to Microsoft 365, you can boost productivity and make better decisions with the information you already own.

External Cloud and API Integration

You can extend your AI agents beyond Microsoft services. Foundry lets you connect to AWS, Google Cloud, and many other platforms. You can use APIs to pull in data from external sources or push results to other systems. This flexibility helps you break down data silos and create a unified view of your business.

Here is a table that shows how Foundry’s integration features support enterprise AI solutions:

Feature	Description
Native Integration with MS365 Ecosystem	Foundry connects seamlessly with MS365, enhancing its utility for enterprise AI solutions.
Automatic Recognition of SharePoint ACLs	This feature simplifies access control management within the integrated environment.
Integration with Microsoft Copilot Studio	Enhances AI capabilities by leveraging Microsoft's AI tools directly within Foundry.
Workflow Automation via Power Automate	Offers over 1,400 connectors for automating workflows, increasing efficiency in enterprise tasks.

You can automate workflows across clouds and on-premises systems. Foundry’s connectors help you link agents to SaaS apps, databases, and internal APIs. This makes it easier to share knowledge and automate business processes across your entire organization.

Extensibility with Azure AI Foundry SDK

You can use the Azure AI Foundry SDK to customize and extend your AI agents. The SDK supports popular frameworks like Semantic Kernel, AutoGen, CrewAI, LangGraph, and LlamaIndex. You can choose the tools that work best for your team. Foundry also supports open protocols such as MCP and A2A. This lets your agents connect with external systems and work together across different platforms.

The SDK gives you built-in tools for rapid development. You can create functional agents quickly and reduce the time it takes to deliver value. Foundry integrates observability and CI/CD practices into your workflow. This ensures your agents stay reliable and safe in production.

Here is a table that highlights how the SDK enhances your AI projects:

Aspect	Description
Support for various frameworks	Use first-party and third-party frameworks to build agents your way.
Interoperability with open protocols	Connect agents to external systems and collaborate across runtimes.
Built-in tools for rapid value	Create functional agents quickly with ready-to-use tools.
Enterprise-grade management	Ensure reliability and safety with integrated observability and CI/CD.

You can also wrap your own business systems as agentic AI tools. This makes them portable and discoverable across teams. Foundry’s connectors reach over 1,400 SaaS and on-premises systems. You can integrate your agents with almost any enterprise process, giving your organization a competitive edge.

Note: Extensibility with the Azure AI Foundry SDK helps you adapt your AI solutions as your business grows and changes.

Real-World Use Cases and Lessons Learned

Automating Business Processes

You can use Microsoft Foundry to automate many business processes in your organization. Foundry lets you build agents that work across time boundaries. These agents manage tasks that may take hours, days, or even weeks. You can orchestrate tools and models to handle complex workflows. Durable agents keep your operations running smoothly, even when tasks span long periods.

Agents help you reduce the time needed to complete business processes.
You can improve productivity by letting agents handle routine tasks.
Agents support better decision-making by providing timely information.
You can lower costs by automating repetitive work.
Agents continuously evaluate and improve workflows.

For example, you can automate document management. Agents can organize files, extract key information, and route documents for approval. You can also retrieve data from multiple clouds, breaking down silos and giving your team access to the information they need. These solutions help you focus on important work and drive business growth.

Tip: Durable, stateful agents ensure your workflows stay on track, even when tasks require ongoing attention.

Overcoming Deployment Challenges

You may face several challenges when deploying AI agents with Microsoft Foundry. You need to understand these obstacles and use proven strategies to overcome them. The table below shows common challenges and effective solutions:

Challenges	Strategies
Generative AI model limitations	Evaluate the model before incorporation to understand its limitations.
Tool orchestration complexities	Choose and integrate tools thoughtfully to ensure stability and proper documentation.
Unequal representation and support	Provide user proactive controls for system boundaries to enhance performance across diverse groups.
Opaque decision-making processes	Ensure intelligibility and traceability for human decision-making to help users understand agent actions.
Evolving best practices and standards	Establish real-time oversight and human-in-the-loop processes for critical tasks.

You can address model limitations by testing and validating before deployment. Careful tool selection and documentation help you build stable systems. You should give users control over agent boundaries to support diverse needs. Clear traceability lets you understand agent decisions. Real-time oversight and human involvement keep your AI solutions safe and effective.

Note: You can overcome deployment challenges by planning ahead and using strategies that fit your organization’s needs.

Best Practices for Production Grade AI Agents

You can follow best practices to ensure your AI agents perform well in production. These practices help you build reliable, secure, and efficient solutions.

Validate all inputs and outputs for every tool.
Use role-based access control with Azure Entra ID and managed identities.
Protect your data with private endpoints, VNet integration, and encryption.
Log every tool invocation and reasoning step to Application Insights.
Monitor tokens, costs, and reasoning depth.
Instrument agents early for full observability.
Document and enforce tool contracts.
Test agents with adversarial prompts.
Set clear observability and governance policies.
Include performance and load testing.
Build a continuous evaluation loop.

You can improve reliability by monitoring agents and logging actions. Security features like encryption and access control keep your data safe. Testing and documentation help you catch problems early. Continuous evaluation lets you adapt and improve your agents as your business changes.

Tip: You can achieve success by combining technical best practices with strong oversight and regular evaluation.

Steps to Build and Deploy with Foundry

Planning and Requirements

You should begin by defining what your AI agent needs to do. Start with a clear scope and outline the agent’s main tasks. Before you write any code, gather the expertise your team has learned and document it. This step helps you avoid confusion later.

Choose a platform that matches your company’s needs for evaluation and governance.
Design your agent’s behavior and give it a clear identity.
Test your ideas with real scenarios. Use acceptance tests to check if the agent can handle important tasks, such as identifying project stages or mapping governance needs.
Expect to make changes. Early versions may not work perfectly, so plan for several rounds of improvement.
Look at how other industries solve similar problems. This can help you spot patterns and avoid mistakes.

The agent passed three acceptance tests. It identified project management stages, mapped governance to high-stakes environments, and recognized influence dynamics. This shows the value of real-world validation.

Start with a pilot project that carries low or medium risk. Monitor everything using tools like OpenTelemetry. Make sure each agent action follows the principle of least privilege. For important decisions, keep a human in the loop. Always check third-party models and data flows before you go live.

Development and Testing

When you build your agent, follow a step-by-step process. First, set up Azure AI Foundry. Then, create and test your prompt flow. Connect any external tools your agent will use. If your agent needs to remember past actions, add memory features.

Use a structured evaluation plan for each agent. Test your agent in different situations to see how it performs. Simulation-based testing lets you compare the agent’s answers to what you expect. Measure results with both numbers and feedback. This helps you see if your agent meets quality standards and covers all needed tasks.

Deployment and Continuous Improvement

After testing, deploy your agent using Azure Bot Service. Keep monitoring your agent’s performance. Use dashboards and alerts to track how well it works.

Feature	Description
Scaling and Productionization	Quickly turn prototypes into production agents. Monitor and refine them as you go.
Enterprise Governance and Compliance	Keep data private and follow regulations with built-in policies and controls.
Unified Control Plane	Manage all your agents from one place. Apply security rules everywhere.

Most organizations see value from GenAI deployment in about 13 months. You can speed this up by using your existing tools and focusing on clear use cases. Always look for ways to improve your agents. Update them as your business needs change. This approach helps you get the most from Microsoft Foundry and keeps your AI solutions strong.

You can build and deploy production-grade AI agents by following best practices with Microsoft Foundry. Hybrid architecture, seamless integration, and strong governance drive real results:

Metric	Value
ROI over three years	327%
Payback period	Less than 6 months
Developer productivity boost	Up to 35%
Enterprise benefits	Nearly $49.5M

No-code tools and SDKs help you move faster:

Deploy agents directly with hosted runtimes.
Design complex workflows visually.
Start small, validate quickly, and scale responsibly.

Organizations in manufacturing, logistics, and healthcare have already improved efficiency and reduced costs with Foundry. Now is the time to explore what you can achieve.

FAQ

What is Microsoft Foundry?

Microsoft Foundry is a platform that helps you build, deploy, and manage AI agents for your business. You can use it to automate tasks, connect to data, and ensure security.

How do you start building an AI agent with Foundry?

You begin by defining your agent’s tasks. Use no-code tools or the SDK to create workflows. Test your agent in the model playground before deploying it.

Can you connect Foundry to other cloud services?

Yes, you can link Foundry to AWS, Google Cloud, and many SaaS apps. This helps you break down data silos and access information across your organization.

How does Foundry keep your data secure?

Foundry uses role-based access control, managed identities, and encryption. You can monitor agent actions and set permissions to protect sensitive information.

What tools help you monitor AI agents in Foundry?

You get dashboards, real-time alerts, and logging tools. These features let you track agent performance, spot issues, and maintain compliance.

Do you need coding skills to use Foundry?

No, you can use no-code tools to build and deploy agents. If you want advanced features, you can use the SDK and integrate custom code.

How does Foundry support responsible AI practices?

Foundry gives you tools for data classification, policy enforcement, and audit logging. You can combine these with human oversight to ensure ethical outcomes.

What industries benefit most from Foundry?

Manufacturing, logistics, healthcare, and finance use Foundry to automate processes, improve efficiency, and reduce costs. You can adapt Foundry to fit your industry’s needs.

🚀 Want to be part of m365.fm?

Then stop just listening… and start showing up.

👉 Connect with me on LinkedIn and let’s make something happen:

🎙️ Be a podcast guest and share your story
🎧 Host your own episode (yes, seriously)
💡 Pitch topics the community actually wants to hear
🌍 Build your personal brand in the Microsoft 365 space

This isn’t just a podcast — it’s a platform for people who take action.

🔥 Most people wait. The best ones don’t.

👉 Connect with me on LinkedIn and send me a message:
"I want in"

Let’s build something awesome 👊

1
00:00:00,000 --> 00:00:07,000
Yeah, welcome to another edition of the M365 at AMP podcast.

2
00:00:07,000 --> 00:00:20,000
My guest is Edgar machine today and he is a data and AI engineer, Microsoft MVP for Microsoft Foundry and Business application based in Nairobi.

3
00:00:20,000 --> 00:00:31,000
And he also run his own labs and I'm interested in it's your community, but can you tell a little bit about your community.

4
00:00:31,000 --> 00:00:34,000
Thank you so much, my crew.

5
00:00:34,000 --> 00:00:42,000
So thank you once again, my name is Edgar, the Microsoft MVP in Microsoft Foundry and Business applications.

6
00:00:42,000 --> 00:00:52,000
So one thing I agree with this mantra like I always love to met people and to update people.

7
00:00:52,000 --> 00:01:00,000
So please with my colleagues, we started our own lab here in Kenya.

8
00:01:00,000 --> 00:01:14,000
So basically we wanted to bring a campus students and I like graduate. So the people who just graduated from campus and they want to improve their skills and this.

9
00:01:14,000 --> 00:01:19,000
This enterprise grade knowledge and to gain.

10
00:01:19,000 --> 00:01:33,000
So we started the my coach England and currently we are learning what you see have so would you see this as what he named for land, land so that would you see.

11
00:01:33,000 --> 00:01:36,000
So would you see how we have it.

12
00:01:36,000 --> 00:01:46,000
So we have a little bit of a little bit of a little bit of a little bit of a little bit of a little bit mostly in Microsoft platform.

13
00:01:46,000 --> 00:01:57,000
So we have the low, low, low, which are the business application apps or the platform where we have our app and pages, copy, studio.

14
00:01:57,000 --> 00:02:11,000
And we have the data platform, which is the Microsoft fabric and we have the now the procured where we have the developed tools and Microsoft foundry as a full procured language.

15
00:02:11,000 --> 00:02:15,000
So we ensure that all of our.

16
00:02:15,000 --> 00:02:17,940
or maintain how access to these applications

17
00:02:17,940 --> 00:02:20,920
so that they can be able to screen, land,

18
00:02:20,920 --> 00:02:22,740
and that's really built on JavaScript.

19
00:02:22,740 --> 00:02:26,960
So that's a brief overview for managing labs.

20
00:02:26,960 --> 00:02:29,600
- Yeah, this is awesome.

21
00:02:29,600 --> 00:02:34,600
So also I say you have to find all the links in the showed notes

22
00:02:34,600 --> 00:02:39,200
so it's as available to look at the page

23
00:02:39,200 --> 00:02:43,440
but today, on topic is building and deploying product-grate,

24
00:02:43,440 --> 00:02:45,960
AI agents with Foundry.

25
00:02:45,960 --> 00:02:50,960
And so can you tell us a little bit about your experience

26
00:02:50,960 --> 00:02:55,200
'cause Microsoft Azure and yeah, the AI related services?

27
00:02:55,200 --> 00:03:01,520
- So I started learning Microsoft product

28
00:03:01,520 --> 00:03:02,880
when I was in campus.

29
00:03:02,880 --> 00:03:05,000
So that was back in 2022.

30
00:03:05,000 --> 00:03:10,000
I was very curious actually to know how I can be able

31
00:03:10,000 --> 00:03:13,200
to build using virtual lead that time.

32
00:03:13,200 --> 00:03:15,600
I was so focused in Azure.

33
00:03:15,600 --> 00:03:18,000
So club computing was, you know,

34
00:03:18,000 --> 00:03:19,720
when you're in campus, everyone tells you

35
00:03:19,720 --> 00:03:23,120
you need to land anything about club computing.

36
00:03:23,120 --> 00:03:28,040
So we had AWS and Azure and I was actually loved Azure.

37
00:03:28,040 --> 00:03:29,840
So I started learning Azure,

38
00:03:29,840 --> 00:03:32,720
but in campus and when I was in,

39
00:03:32,720 --> 00:03:37,480
that's now when I started the Microsoft project.

40
00:03:37,480 --> 00:03:42,480
So I have experience both in just-and-a-i.

41
00:03:43,160 --> 00:03:48,160
So I can say like, my piece of experience

42
00:03:48,160 --> 00:03:51,520
and built applications.

43
00:03:51,520 --> 00:03:56,520
So this application, basically we have the African staff

44
00:03:56,520 --> 00:04:02,320
where we create African products.

45
00:04:02,320 --> 00:04:06,640
So for example, we have the e-commerce Kenya

46
00:04:06,640 --> 00:04:11,320
has a very, very, very large e-commerce system where

47
00:04:11,320 --> 00:04:16,320
we have more business people who want to create the e-commerce

48
00:04:16,320 --> 00:04:20,920
the e-commerce website, the e-commerce platform

49
00:04:20,920 --> 00:04:23,800
and we always use our mobile banking,

50
00:04:23,800 --> 00:04:24,720
which is in-person.

51
00:04:24,720 --> 00:04:27,920
So building up and integrating with our employees

52
00:04:27,920 --> 00:04:31,440
that they measure up things that most business

53
00:04:31,440 --> 00:04:34,880
business is for African Kenya.

54
00:04:34,880 --> 00:04:39,400
So apart from that, I also build worldwide,

55
00:04:40,920 --> 00:04:42,560
worldwide application.

56
00:04:42,560 --> 00:04:45,400
So I'm from, I'm also, I'm a,

57
00:04:45,400 --> 00:04:49,400
it's an organization that is built in America

58
00:04:49,400 --> 00:04:50,640
or remotely.

59
00:04:50,640 --> 00:04:55,320
So I've been able to create applications for clients,

60
00:04:55,320 --> 00:04:58,240
basically, applications.

61
00:04:58,240 --> 00:05:02,360
Yeah, so that's has been my experience.

62
00:05:02,360 --> 00:05:05,480
Yeah.

63
00:05:05,480 --> 00:05:10,680
- Can you tell a little bit more about the AI and generative?

64
00:05:10,680 --> 00:05:13,800
I think you have implemented in those platforms.

65
00:05:13,800 --> 00:05:19,920
- Yeah, so the biggest project that happened in the last year,

66
00:05:19,920 --> 00:05:24,960
so at family, we built the Mela,

67
00:05:24,960 --> 00:05:29,160
I Mela is our internal application.

68
00:05:29,160 --> 00:05:31,400
So the idea behind Mela,

69
00:05:31,400 --> 00:05:36,400
I was to create an iSystem that is enterprise

70
00:05:36,400 --> 00:05:39,600
and suited for a particular organization.

71
00:05:39,600 --> 00:05:44,600
So it can be able to know everything that Chorganization does,

72
00:05:44,600 --> 00:05:48,640
it is connected to your enterprise knowledge.

73
00:05:48,640 --> 00:05:52,320
So for example, if you're a Microsoft Fata

74
00:05:52,320 --> 00:05:54,680
or you use Microsoft ecosystem,

75
00:05:54,680 --> 00:05:58,400
so our Mela connects to SharePoint,

76
00:05:58,400 --> 00:06:01,080
you can be able to pull data from SharePoint,

77
00:06:01,080 --> 00:06:04,080
you can be able to use Nugra connectors,

78
00:06:04,080 --> 00:06:08,320
you can be able to now connect to the Outflow,

79
00:06:08,320 --> 00:06:10,080
you can connect to your,

80
00:06:10,080 --> 00:06:14,640
the Outflow connect to one drive connect to,

81
00:06:14,640 --> 00:06:17,800
yeah, basically any information that you have

82
00:06:17,800 --> 00:06:19,840
in your Microsoft ecosystem,

83
00:06:19,840 --> 00:06:22,080
Mela can be able to retrieve them.

84
00:06:22,080 --> 00:06:26,240
Apart from that, we connect using anything that now

85
00:06:26,240 --> 00:06:28,480
is beyond Microsoft.

86
00:06:28,480 --> 00:06:33,480
So let's say you have your data in the Google Cloud of AWS,

87
00:06:33,480 --> 00:06:36,480
can be able to pull them in.

88
00:06:36,480 --> 00:06:39,760
So this can be what we do for Mela,

89
00:06:39,760 --> 00:06:41,560
it's a backup application.

90
00:06:41,560 --> 00:06:45,880
So it can be able to help you find data whenever you,

91
00:06:45,880 --> 00:06:48,960
let's say most organization actually have data

92
00:06:48,960 --> 00:06:51,120
that is cut everywhere.

93
00:06:51,120 --> 00:06:55,400
So having these applications that you can just query

94
00:06:55,400 --> 00:06:58,800
and it can give you the direct location

95
00:06:58,800 --> 00:07:01,080
where that document is found,

96
00:07:01,080 --> 00:07:03,800
you can basically ask any question

97
00:07:03,800 --> 00:07:06,800
about show organization, the procedures,

98
00:07:06,800 --> 00:07:09,440
you can request for links and everything.

99
00:07:09,440 --> 00:07:13,400
So that's Mela and I've told that

100
00:07:13,400 --> 00:07:16,880
Mela have a lot of capabilities where you can be able to

101
00:07:16,880 --> 00:07:22,880
bring local models or cloud models.

102
00:07:22,880 --> 00:07:25,680
So for the past year, we've noticed

103
00:07:25,680 --> 00:07:31,400
that the course of AI has been actually going to get,

104
00:07:32,800 --> 00:07:35,800
is something that truth experience personally

105
00:07:35,800 --> 00:07:37,760
and I have experience with.

106
00:07:37,760 --> 00:07:41,760
So when the cloud I came,

107
00:07:41,760 --> 00:07:45,640
I actually, we actually connected a cloud

108
00:07:45,640 --> 00:07:50,160
as one of the LLM that we use in Mela AI.

109
00:07:50,160 --> 00:07:55,160
So apparently when we had the API and LLM

110
00:07:55,160 --> 00:08:00,800
connected to Mela, most of our employees were using

111
00:08:00,800 --> 00:08:03,360
cloud and our course for cloud actually,

112
00:08:03,360 --> 00:08:06,760
that one it actually started a lot.

113
00:08:06,760 --> 00:08:10,480
So it was a lesson that I actually learned

114
00:08:10,480 --> 00:08:14,360
and our organization is it's not bad way,

115
00:08:14,360 --> 00:08:18,640
but we had to, you know, mistakes mix you,

116
00:08:18,640 --> 00:08:21,800
land and it helps you solve them.

117
00:08:21,800 --> 00:08:24,720
So how we try to solve this whenever

118
00:08:24,720 --> 00:08:29,240
these cloud models get the high course

119
00:08:29,240 --> 00:08:31,800
because you know, the global model

120
00:08:31,800 --> 00:08:35,360
actually the course can be quite high.

121
00:08:35,360 --> 00:08:38,800
So we created a local model.

122
00:08:38,800 --> 00:08:42,120
So this local model can be run,

123
00:08:42,120 --> 00:08:45,600
either locally in your own PC

124
00:08:45,600 --> 00:08:48,560
or we can let it run in cloud.

125
00:08:48,560 --> 00:08:51,080
So in our, we have the phone,

126
00:08:51,080 --> 00:08:54,000
the local, they only first take agents where

127
00:08:54,000 --> 00:08:56,120
we can be able to post a agent

128
00:08:56,120 --> 00:08:57,840
to run the liquid cloud,

129
00:08:57,840 --> 00:09:01,640
we train them and we now expose the input.

130
00:09:01,640 --> 00:09:03,720
So when this endpoint reconnects,

131
00:09:03,720 --> 00:09:07,160
but it's not Mela and it will be able to use

132
00:09:07,160 --> 00:09:10,400
the LLM pretty much for free.

133
00:09:10,400 --> 00:09:12,840
So let us be my experience.

134
00:09:12,840 --> 00:09:16,600
So we have the ability where you can be able to create

135
00:09:16,600 --> 00:09:20,160
a local model connected to our application.

136
00:09:20,160 --> 00:09:23,600
No, it becomes hybrid and hybrid,

137
00:09:23,600 --> 00:09:26,600
honestly, I felt like hybrid has been

138
00:09:26,600 --> 00:09:30,080
the biggest change in Mela AI

139
00:09:30,080 --> 00:09:32,480
where we have the ability to switch.

140
00:09:32,480 --> 00:09:36,720
So we have that orchestration layer

141
00:09:36,720 --> 00:09:40,440
where whenever someone has a question,

142
00:09:40,440 --> 00:09:43,400
our AI actually chooses which modern

143
00:09:43,400 --> 00:09:45,320
is the best to respond to.

144
00:09:45,320 --> 00:09:47,320
So if the work is not complex,

145
00:09:47,320 --> 00:09:49,640
it will automatically switch to local models

146
00:09:49,640 --> 00:09:52,360
so that it's the cost zone.

147
00:09:52,360 --> 00:09:56,320
And when we have something so complex

148
00:09:56,320 --> 00:10:00,320
that our employees want to actually do

149
00:10:00,320 --> 00:10:03,040
it knows which is to the cloud model.

150
00:10:03,040 --> 00:10:07,360
So let us make a experience using the AI.

151
00:10:07,360 --> 00:10:11,920
- Oh, that's really interesting.

152
00:10:11,920 --> 00:10:14,920
Yesterday I also talked about pricing.

153
00:10:14,920 --> 00:10:19,400
Can you say how big was it decreased in pricing

154
00:10:19,400 --> 00:10:21,400
for the project from an AI?

155
00:10:22,960 --> 00:10:27,960
Actually, it increased drastically because of it was

156
00:10:27,960 --> 00:10:31,800
actually quite expensive.

157
00:10:31,800 --> 00:10:36,800
The first three months where we connected the anthropic models,

158
00:10:36,800 --> 00:10:41,640
we had we had a who's and sonnet connected.

159
00:10:41,640 --> 00:10:46,640
So whenever our employees were using our Mela AI,

160
00:10:46,640 --> 00:10:50,400
they all looked up to these anthropic models.

161
00:10:50,400 --> 00:10:55,000
So the first month was just protesting because now

162
00:10:55,000 --> 00:10:59,160
we opened it up to a test group.

163
00:10:59,160 --> 00:11:02,920
So we have a department for just purely testing.

164
00:11:02,920 --> 00:11:07,600
So they went on and as you say in handling,

165
00:11:07,600 --> 00:11:09,600
we hit our whole production.

166
00:11:09,600 --> 00:11:14,080
So they went, broke it, tried it, tested everything

167
00:11:14,080 --> 00:11:16,600
just to make sure that people's working very well.

168
00:11:16,600 --> 00:11:19,400
So what ended up happening is

169
00:11:19,400 --> 00:11:22,000
they were going to be satisfied.

170
00:11:22,000 --> 00:11:26,040
So when we noticed that course actually was high

171
00:11:26,040 --> 00:11:29,240
and maintaining that course for one year,

172
00:11:29,240 --> 00:11:31,320
actually it was so damaging.

173
00:11:31,320 --> 00:11:33,920
So they think we think we actually do.

174
00:11:33,920 --> 00:11:36,200
So the first thing we did was actually

175
00:11:36,200 --> 00:11:38,200
I think a hybrid model.

176
00:11:38,200 --> 00:11:42,480
So a hybrid model is where we now connect

177
00:11:42,480 --> 00:11:45,400
and local models to under cloud models.

178
00:11:45,400 --> 00:11:47,120
So it's vicious.

179
00:11:47,120 --> 00:11:51,320
So whenever our task is low, use the local model,

180
00:11:51,320 --> 00:11:54,640
whenever we have a task that is much more complex

181
00:11:54,640 --> 00:11:57,360
than switches to the cloud model.

182
00:11:57,360 --> 00:12:01,400
And the second now, who's now really limiting.

183
00:12:01,400 --> 00:12:06,000
And this is very important and really limiting

184
00:12:06,000 --> 00:12:08,120
in AI applications.

185
00:12:08,120 --> 00:12:14,520
So we have the ability to govern the usage of the AI model.

186
00:12:14,520 --> 00:12:17,720
So we have the limiting in cloud models.

187
00:12:17,720 --> 00:12:21,240
So it can be in either the GPT models there

188
00:12:21,240 --> 00:12:23,960
and the GPT models or the GTI model.

189
00:12:23,960 --> 00:12:28,640
So we can be able to monitor that, they will say.

190
00:12:28,640 --> 00:12:31,920
So whenever I use that, I use such spike,

191
00:12:31,920 --> 00:12:36,560
I will throttle the usage of that application.

192
00:12:36,560 --> 00:12:42,240
So that you will mostly find that that special will actually

193
00:12:42,240 --> 00:12:42,920
expire.

194
00:12:42,920 --> 00:12:47,080
And there will not notice each time because it will now switch

195
00:12:47,080 --> 00:12:49,040
to a lower model.

196
00:12:49,040 --> 00:12:55,520
So that's a lot because now we are able now to put a budget.

197
00:12:55,520 --> 00:12:59,400
So now this is step back and side.

198
00:12:59,400 --> 00:13:04,280
Each user has a budget for each model that they use.

199
00:13:04,280 --> 00:13:08,160
So let's say for example, we are using their anthropic models.

200
00:13:08,160 --> 00:13:12,600
And we have a budget for a user, let's say,

201
00:13:12,600 --> 00:13:14,040
$10 a month.

202
00:13:14,040 --> 00:13:19,360
So they cannot take it that $10.

203
00:13:19,360 --> 00:13:22,920
So whenever they get $10, it will get switched

204
00:13:22,920 --> 00:13:27,480
and they will use the GPT models.

205
00:13:27,480 --> 00:13:28,640
OK.

206
00:13:28,640 --> 00:13:29,960
This sounds awesome.

207
00:13:29,960 --> 00:13:30,440
Thank you.

208
00:13:30,440 --> 00:13:31,440
Thank you.

209
00:13:31,440 --> 00:13:32,640
OK.

210
00:13:32,640 --> 00:13:37,480
Let's a little bit jump more into Azure AI Foundry Fundamentals.

211
00:13:37,480 --> 00:13:42,320
So what is Microsoft AI Foundry?

212
00:13:42,320 --> 00:13:47,800
And what's the business problems do it solve?

213
00:13:47,800 --> 00:13:54,280
So in Microsoft Foundry, this is how I can explain it.

214
00:13:54,280 --> 00:13:57,040
I can't say that there is a playground.

215
00:13:57,040 --> 00:14:01,120
So whenever you want to play with this application

216
00:14:01,120 --> 00:14:05,720
to play with the models, it's a platform where

217
00:14:05,720 --> 00:14:11,720
you can create AI application data in a local environment.

218
00:14:11,720 --> 00:14:15,920
In a local manner or fully-procored manner.

219
00:14:15,920 --> 00:14:23,280
So it has a lot of capabilities when it's a full infrastructure

220
00:14:23,280 --> 00:14:28,320
that you can be able to actually get any model that you want.

221
00:14:28,320 --> 00:14:34,280
So it has the anthropic models, it has the GPT models,

222
00:14:34,280 --> 00:14:38,360
the mistress of the modern market base.

223
00:14:38,360 --> 00:14:41,360
So whenever you want to test a particular model,

224
00:14:41,360 --> 00:14:44,840
you can always work there such for a particular model.

225
00:14:44,840 --> 00:14:47,120
And actually, it has a list.

226
00:14:47,120 --> 00:14:53,280
So you can see the model that is performing the pairs of--

227
00:14:53,280 --> 00:14:57,560
the first, let me talk about the features of Microsoft Foundry.

228
00:14:57,560 --> 00:15:02,160
In my explanation, Microsoft Foundry is just a model playground.

229
00:15:02,160 --> 00:15:05,280
If you want to create business applications

230
00:15:05,280 --> 00:15:11,040
for any AI application, either in local manner or in a pro-code

231
00:15:11,040 --> 00:15:15,120
model, you can be able to create it, you can be able to host it.

232
00:15:15,120 --> 00:15:20,080
And within a lot of the most important Microsoft Foundry,

233
00:15:20,080 --> 00:15:24,440
is you have the ability to connect with enterprise connectors.

234
00:15:24,440 --> 00:15:27,320
So you can-- it has the--

235
00:15:27,320 --> 00:15:32,120
honestly, I found that Foundry has the best things.

236
00:15:32,120 --> 00:15:37,520
He's from my experience, Foundry has the best MCP.

237
00:15:37,520 --> 00:15:43,280
So this helps to connect to your whole ecosystem.

238
00:15:43,280 --> 00:15:49,080
So if you have data that is working at a SharePoint 1 drive,

239
00:15:49,080 --> 00:15:52,320
how to look, you can be able to bring them in.

240
00:15:52,320 --> 00:15:55,760
So they can eat the work IQ.

241
00:15:55,760 --> 00:15:58,960
So you can be able to bring in all the enterprise

242
00:15:58,960 --> 00:16:01,240
or the structure, but apart from that,

243
00:16:01,240 --> 00:16:04,400
it also has the enterprise connectors,

244
00:16:04,400 --> 00:16:09,440
where you can be able to connect to any data that is outside,

245
00:16:09,440 --> 00:16:13,160
now, the Microsoft ecosystem.

246
00:16:13,160 --> 00:16:17,800
So that's one feature that I actually left the most.

247
00:16:17,800 --> 00:16:21,920
So apart from that, the next thing that Foundry has

248
00:16:21,920 --> 00:16:28,480
is their how can they say their LLM evaluation.

249
00:16:28,480 --> 00:16:32,840
So you can have the safety evaluation

250
00:16:32,840 --> 00:16:36,720
where you can control how that LLM is used.

251
00:16:36,720 --> 00:16:42,600
So in certain evaluation, not be that the first--

252
00:16:42,600 --> 00:16:48,520
the first principle of AI, where we need AI to be transparent,

253
00:16:48,520 --> 00:16:53,000
we need it to be safer so that the evaluation that

254
00:16:53,000 --> 00:16:54,600
Microsoft Foundry has.

255
00:16:54,600 --> 00:16:58,920
So you can use the questions to control how

256
00:16:58,920 --> 00:17:05,400
you can be able to use the LLM that you have to Foundry.

257
00:17:05,400 --> 00:17:11,240
Apart from that, you can be able to tune your models.

258
00:17:11,240 --> 00:17:15,720
You can be able to create an agentic orchestration.

259
00:17:15,720 --> 00:17:18,560
So an objectic orchestration is--

260
00:17:18,560 --> 00:17:22,880
let's say, for example, you have a lot--

261
00:17:22,880 --> 00:17:26,360
you want to build maybe an ERP system.

262
00:17:26,360 --> 00:17:31,400
And in that ERP system, it has a lot of--

263
00:17:31,400 --> 00:17:34,880
and ERP system is a complex system.

264
00:17:34,880 --> 00:17:38,560
So you want different capabilities.

265
00:17:38,560 --> 00:17:42,240
So you can create different models or different agents

266
00:17:42,240 --> 00:17:44,280
to perform a thing with that.

267
00:17:44,280 --> 00:17:47,080
Then you have one agent.

268
00:17:47,080 --> 00:17:51,480
So we can leave the orchestrator agent that can now call

269
00:17:51,480 --> 00:17:55,760
the different worker agent so that they can work with together.

270
00:17:56,640 --> 00:17:59,960
So what does it mean?

271
00:17:59,960 --> 00:18:02,080
Actually, if you leave me to talk about,

272
00:18:02,080 --> 00:18:05,120
I just don't call this.

273
00:18:05,120 --> 00:18:07,120
Yeah.

274
00:18:07,120 --> 00:18:09,520
Please tell me if I call--

275
00:18:09,520 --> 00:18:12,200
and most people say that I am always

276
00:18:12,200 --> 00:18:12,960
too technical.

277
00:18:12,960 --> 00:18:16,240
So whenever you feel I'm technical, you can--

278
00:18:16,240 --> 00:18:17,520
yeah, feel free.

279
00:18:17,520 --> 00:18:19,160
I like it.

280
00:18:19,160 --> 00:18:19,680
I like it.

281
00:18:19,680 --> 00:18:21,720
No, very interesting.

282
00:18:21,720 --> 00:18:24,920
I last have seen another product.

283
00:18:24,920 --> 00:18:27,560
And I try a little bit to different.

284
00:18:27,560 --> 00:18:29,160
There's an ERP.

285
00:18:29,160 --> 00:18:30,160
I found it.

286
00:18:30,160 --> 00:18:31,560
And there is also an ERP.

287
00:18:31,560 --> 00:18:32,960
I study you.

288
00:18:32,960 --> 00:18:36,840
Can you explain what's different?

289
00:18:36,840 --> 00:18:41,040
For its performance to deal.

290
00:18:41,040 --> 00:18:42,880
That's OK.

291
00:18:42,880 --> 00:18:49,920
So there are three products that make

292
00:18:49,920 --> 00:18:51,400
a crossover.

293
00:18:51,400 --> 00:18:55,040
So we have the three ways you can be able to build agent

294
00:18:55,040 --> 00:18:57,520
or skin microsoft.

295
00:18:57,520 --> 00:19:01,080
So the first one is the M365.

296
00:19:01,080 --> 00:19:05,000
So M365, how I can explain it, it's like--

297
00:19:05,000 --> 00:19:09,920
you see the chat GPT data there, the cloud data,

298
00:19:09,920 --> 00:19:14,760
that chat data, that's how the M365 is like.

299
00:19:14,760 --> 00:19:16,760
So it's a couple of times where you can

300
00:19:16,760 --> 00:19:20,960
be able to create agents in--

301
00:19:20,960 --> 00:19:22,880
you can just give a prompt.

302
00:19:22,880 --> 00:19:24,920
I wish I could be sharing my screen.

303
00:19:24,920 --> 00:19:29,360
So you can just give a prompt and it

304
00:19:29,360 --> 00:19:31,360
can create that agent for you.

305
00:19:31,360 --> 00:19:32,840
So that's the first level.

306
00:19:32,840 --> 00:19:36,520
So the first level is the simplest level where you just

307
00:19:36,520 --> 00:19:39,720
give a prompt and it can be able to create an agent.

308
00:19:39,720 --> 00:19:43,400
So after that, you go to step two.

309
00:19:43,400 --> 00:19:45,600
Step two now is the co-finals studio.

310
00:19:45,600 --> 00:19:49,560
It's a local and settlement where

311
00:19:49,560 --> 00:19:54,200
you can be able to create agents in no code.

312
00:19:54,200 --> 00:19:56,160
It's no code and no code.

313
00:19:56,160 --> 00:20:01,800
So no code is just if it is structured between go ahead

314
00:20:01,800 --> 00:20:05,640
and create that project for you or that agent for you.

315
00:20:05,640 --> 00:20:08,840
And if you want to do the procode,

316
00:20:08,840 --> 00:20:12,400
you can be able to actually add in your own code.

317
00:20:12,400 --> 00:20:14,400
So that's the co-finals studio.

318
00:20:14,400 --> 00:20:16,280
So co-finals studio can be able also

319
00:20:16,280 --> 00:20:19,400
to connect with other enterprise applications,

320
00:20:19,400 --> 00:20:21,320
another enterprise connectors.

321
00:20:21,320 --> 00:20:25,520
So I do this in my labs.

322
00:20:25,520 --> 00:20:29,360
So mostly most of my students and new

323
00:20:29,360 --> 00:20:32,160
to the Microsoft ecosystem.

324
00:20:32,160 --> 00:20:37,160
So basically last month, we majorly touched on our platform.

325
00:20:37,160 --> 00:20:41,800
So that was the place where co-finals studio and pound platform.

326
00:20:41,800 --> 00:20:45,400
So now when you're not fully confident enough,

327
00:20:45,400 --> 00:20:49,560
this is where most people, if are the use of endless

328
00:20:49,560 --> 00:20:51,640
studio and pound.

329
00:20:51,640 --> 00:20:54,520
So co-finals studio is the second set set set,

330
00:20:54,520 --> 00:20:57,080
but you are confined.

331
00:20:57,080 --> 00:21:01,160
It's like you are put in a container.

332
00:21:01,160 --> 00:21:03,520
You can build anything instead of that container,

333
00:21:03,520 --> 00:21:06,760
but now when you want to stretch out of the container,

334
00:21:06,760 --> 00:21:09,760
you can't because everything, for me, our effort,

335
00:21:09,760 --> 00:21:11,560
everything was contained.

336
00:21:11,560 --> 00:21:14,440
So now when you want to be in the building,

337
00:21:14,440 --> 00:21:18,160
anything that you want, like your imagination,

338
00:21:18,160 --> 00:21:21,480
that's where we go to Microsoft pound.

339
00:21:21,480 --> 00:21:25,000
So in Microsoft pound, this is where you can build anything.

340
00:21:25,000 --> 00:21:26,280
You're a man of genius.

341
00:21:26,280 --> 00:21:28,960
Whatever you want, you can bring it in.

342
00:21:28,960 --> 00:21:31,920
So that's the biggest difference.

343
00:21:31,920 --> 00:21:35,000
So the biggest difference between now co-finals studio,

344
00:21:35,000 --> 00:21:40,000
that's the studio that Microsoft pound is from my experience,

345
00:21:41,480 --> 00:21:43,920
the whole thing is from the studio, it's the best,

346
00:21:43,920 --> 00:21:47,080
but you work within a confined space.

347
00:21:47,080 --> 00:21:49,240
So when we move to all three now,

348
00:21:49,240 --> 00:21:51,720
that's now you have the whole one,

349
00:21:51,720 --> 00:21:53,880
so you know you can feel like that, yeah?

350
00:21:53,880 --> 00:21:55,920
You are controlling the world.

351
00:21:55,920 --> 00:21:59,720
So you can build any application, whether you want to take in,

352
00:21:59,720 --> 00:22:04,280
you are in low-code, or you want to bring your whole code,

353
00:22:04,280 --> 00:22:07,960
you want to create an application locally

354
00:22:07,960 --> 00:22:12,960
choosing containerization that was the Foundry for Fed agents,

355
00:22:12,960 --> 00:22:15,960
so you can be able to cause your agents

356
00:22:15,960 --> 00:22:18,360
using now Microsoft pound.

357
00:22:18,360 --> 00:22:22,680
- And when you think on error foundry,

358
00:22:22,680 --> 00:22:26,160
what are the core components of this, I say too old.

359
00:22:26,160 --> 00:22:33,280
- So for me, I feel governance is there.

360
00:22:35,040 --> 00:22:40,040
But that's because now we need to, as I develop,

361
00:22:40,040 --> 00:22:43,480
I need to monitor the usage of,

362
00:22:43,480 --> 00:22:47,080
faster I need to monitor the usage of my application.

363
00:22:47,080 --> 00:22:50,120
I need to see that actually people are not

364
00:22:50,120 --> 00:22:51,680
using that application.

365
00:22:51,680 --> 00:22:56,680
So whether it works within like for an organization,

366
00:22:56,680 --> 00:23:02,040
let it work for that organization,

367
00:23:02,040 --> 00:23:07,040
and keeping confidentiality transparency.

368
00:23:07,040 --> 00:23:11,040
So for me, I feel governance is the most important.

369
00:23:11,040 --> 00:23:15,880
- Okay, and I have seen this, I call it,

370
00:23:15,880 --> 00:23:17,920
prompt flow can you explain this?

371
00:23:17,920 --> 00:23:22,920
- No, but, prompt flow, this is where,

372
00:23:22,920 --> 00:23:27,880
so I can explain prompt flow.

373
00:23:27,880 --> 00:23:32,880
So from from this is where you are able to build

374
00:23:32,880 --> 00:23:37,520
an application within a prompt, so we have to use

375
00:23:37,520 --> 00:23:43,680
to build application in Microsoft Foundry.

376
00:23:43,680 --> 00:23:48,000
The first way is now, we have the Microsoft Foundry

377
00:23:48,000 --> 00:23:49,800
that's at the front floor,

378
00:23:49,800 --> 00:23:52,240
and we have the hosted agents.

379
00:23:52,240 --> 00:23:56,320
So in hosted agents, this is where you bring,

380
00:23:56,320 --> 00:23:59,320
you can build an application locally

381
00:23:59,320 --> 00:24:03,120
and now upload it to Azure and it will be hosted.

382
00:24:03,120 --> 00:24:05,080
So when we talk about prompt flow,

383
00:24:05,080 --> 00:24:08,080
this is where you create an agent

384
00:24:08,080 --> 00:24:10,560
give it your own instructions,

385
00:24:10,560 --> 00:24:13,680
connect to different,

386
00:24:13,680 --> 00:24:17,280
different, okay, say,

387
00:24:17,280 --> 00:24:21,440
you have different connectors,

388
00:24:21,440 --> 00:24:26,960
match and connect to that application.

389
00:24:26,960 --> 00:24:34,240
You have, on the start, you have a little bit

390
00:24:34,240 --> 00:24:36,240
tell about different models.

391
00:24:36,240 --> 00:24:41,800
Is this the model catalog inside the AI Foundry,

392
00:24:41,800 --> 00:24:45,560
or actually,

393
00:24:49,440 --> 00:24:51,320
sorry, sorry, sorry, sorry.

394
00:24:51,320 --> 00:24:52,920
Yeah, I think a little bit,

395
00:24:52,920 --> 00:24:54,920
and that's just this model catalog,

396
00:24:54,920 --> 00:24:56,880
and you have, on the start, you have tell,

397
00:24:56,880 --> 00:25:01,880
you have different models used on your system.

398
00:25:01,880 --> 00:25:07,400
So this is what you use for it inside.

399
00:25:07,400 --> 00:25:11,400
Yeah, so in form, do we have the model catalog?

400
00:25:11,400 --> 00:25:14,400
So in model catalog, this is where it is,

401
00:25:14,400 --> 00:25:16,880
and this is all the

402
00:25:18,760 --> 00:25:22,040
LLM store, any LLM from a flow page,

403
00:25:22,040 --> 00:25:24,080
you pass this on it,

404
00:25:24,080 --> 00:25:25,880
high-cool of course models,

405
00:25:25,880 --> 00:25:28,320
it has the GPT models,

406
00:25:28,320 --> 00:25:30,240
it has the LAMM, the MISROS,

407
00:25:30,240 --> 00:25:34,480
so what model market is,

408
00:25:34,480 --> 00:25:36,320
it's like a marketplace,

409
00:25:36,320 --> 00:25:39,840
where you can be able to see at least for any model

410
00:25:39,840 --> 00:25:41,520
that you want,

411
00:25:41,520 --> 00:25:43,480
you can see how they are performing.

412
00:25:43,480 --> 00:25:45,520
So they are always shredded.

413
00:25:45,520 --> 00:25:46,960
So for example, this week,

414
00:25:46,960 --> 00:25:51,960
you may find that codec is the most performing LLM.

415
00:25:51,960 --> 00:25:54,160
So people actually do it,

416
00:25:54,160 --> 00:25:59,160
so you can be able to see the what codec is best in this week.

417
00:25:59,160 --> 00:26:07,280
They aerate how is it performing that performance spread?

418
00:26:07,280 --> 00:26:11,800
Yeah, so it helps to give you access

419
00:26:11,800 --> 00:26:14,440
or understanding how it is performing,

420
00:26:14,440 --> 00:26:16,840
how you can actually bring it

421
00:26:16,840 --> 00:26:19,600
and from that model catalog,

422
00:26:19,600 --> 00:26:24,600
you can actually purchase those models that you want

423
00:26:24,600 --> 00:26:28,440
and now make them run for your application.

424
00:26:28,440 --> 00:26:33,960
And there are also these evaluation pipelines,

425
00:26:33,960 --> 00:26:37,120
can you tell a little about the performance of it?

426
00:26:37,120 --> 00:26:40,080
Yeah, so in evaluation,

427
00:26:40,080 --> 00:26:45,080
this is now how you test these LLM store.

428
00:26:45,800 --> 00:26:49,640
Let's say for example, I have created an application

429
00:26:49,640 --> 00:26:51,800
that is supposed to start

430
00:26:51,800 --> 00:26:55,880
let anti-combat sites or anti-combat platforms.

431
00:26:55,880 --> 00:26:59,880
So in evaluation, now I can put at least

432
00:26:59,880 --> 00:27:03,160
or maybe drones and list of questions

433
00:27:03,160 --> 00:27:05,080
and see how it performed.

434
00:27:05,080 --> 00:27:10,080
So I can say if a list of expected questions

435
00:27:10,080 --> 00:27:14,680
and expected results,

436
00:27:14,680 --> 00:27:18,840
so it will help me to actually understand

437
00:27:18,840 --> 00:27:22,880
if actually meeting my expectation

438
00:27:22,880 --> 00:27:26,440
and say for example, the evaluation is at 80%,

439
00:27:26,440 --> 00:27:28,960
which moves like it has answered

440
00:27:28,960 --> 00:27:33,440
most of the things that the application I build.

441
00:27:33,440 --> 00:27:36,840
So first, the evaluation has passed

442
00:27:36,840 --> 00:27:39,920
to really test these models.

443
00:27:39,920 --> 00:27:43,080
We can evaluate if it can be able,

444
00:27:43,080 --> 00:27:47,600
this is now in from injection to see if someone can use it

445
00:27:47,600 --> 00:27:52,600
for malicious drones or malicious things that they want.

446
00:27:52,600 --> 00:27:58,040
So the next thing you can evaluate the performance,

447
00:27:58,040 --> 00:28:00,080
how you do really perform it,

448
00:28:00,080 --> 00:28:04,480
you can evaluate the hallucination.

449
00:28:04,480 --> 00:28:07,240
So is it accurate enough?

450
00:28:07,240 --> 00:28:12,240
Is the response you wanted the thing that you expected?

451
00:28:12,240 --> 00:28:14,800
So the evaluation actually helps us

452
00:28:14,800 --> 00:28:18,600
to now actually monitor and see how it actually

453
00:28:18,600 --> 00:28:21,800
performing based on their expected outcome.

454
00:28:21,800 --> 00:28:24,240
- Cool.

455
00:28:24,240 --> 00:28:29,760
I think it's, as I found it, it's an enterprise solution.

456
00:28:29,760 --> 00:28:32,280
So I think there are more roles,

457
00:28:32,280 --> 00:28:35,680
like the developer, data scientists, security teams,

458
00:28:35,680 --> 00:28:38,120
how did you manage collaboration

459
00:28:39,400 --> 00:28:42,480
in Azure Foundry?

460
00:28:42,480 --> 00:28:48,000
- We do it actually.

461
00:28:48,000 --> 00:28:51,560
We have different developers in our team.

462
00:28:51,560 --> 00:28:56,560
So what we do, we use, okay say,

463
00:28:56,560 --> 00:29:05,800
I forgot the best way to explain this.

464
00:29:06,800 --> 00:29:08,560
(laughs)

465
00:29:08,560 --> 00:29:09,640
- Yeah.

466
00:29:09,640 --> 00:29:12,120
But you can be able to collaborate.

467
00:29:12,120 --> 00:29:15,920
So how we do it, it only is,

468
00:29:15,920 --> 00:29:17,640
we create projects in front of you.

469
00:29:17,640 --> 00:29:19,520
So we have different projects.

470
00:29:19,520 --> 00:29:22,320
But the first thing we do is you create a project

471
00:29:22,320 --> 00:29:23,520
and inside that project,

472
00:29:23,520 --> 00:29:28,520
you can be able now to now create agents in that project.

473
00:29:28,520 --> 00:29:33,520
So we have access to different employees

474
00:29:33,520 --> 00:29:36,040
and access to different platforms.

475
00:29:36,040 --> 00:29:37,920
So let's say for example,

476
00:29:37,920 --> 00:29:41,000
I was working for a particular agent.

477
00:29:41,000 --> 00:29:44,280
So we have the opportunity to collaborate

478
00:29:44,280 --> 00:29:46,160
on that particular agent.

479
00:29:46,160 --> 00:29:48,480
Then after it's complete,

480
00:29:48,480 --> 00:29:51,760
we have our testing and dependency team actually

481
00:29:51,760 --> 00:29:53,280
who can do for the testing

482
00:29:53,280 --> 00:29:54,680
if that's a recommendation.

483
00:29:54,680 --> 00:29:57,680
So I think that's how we do it.

484
00:29:57,680 --> 00:30:02,040
And that's how we collaborate inside the Microsoft Form.

485
00:30:03,040 --> 00:30:03,880
- Awesome.

486
00:30:03,880 --> 00:30:07,600
Can you a little bit explain the lifecycle of an AI application

487
00:30:07,600 --> 00:30:09,160
in Azure Foundry?

488
00:30:09,160 --> 00:30:12,600
- So the lifecycle.

489
00:30:12,600 --> 00:30:16,000
So how do we do this?

490
00:30:16,000 --> 00:30:19,360
So the first thing is for the research.

491
00:30:19,360 --> 00:30:22,440
I'm very, very good in research.

492
00:30:22,440 --> 00:30:24,640
So whenever you have an idea,

493
00:30:24,640 --> 00:30:27,000
always thought with the research.

494
00:30:27,000 --> 00:30:30,680
And now an experience in YTCs.

495
00:30:31,560 --> 00:30:35,400
The best thing that I use in Foundry.

496
00:30:35,400 --> 00:30:40,400
So we have VS Code extension in,

497
00:30:40,400 --> 00:30:45,680
yeah, we can be able to connect to from using VS Code.

498
00:30:45,680 --> 00:30:48,680
So the first thing is whenever I have an idea,

499
00:30:48,680 --> 00:30:51,360
I do a properly search on it.

500
00:30:51,360 --> 00:30:53,000
So after doing a research,

501
00:30:53,000 --> 00:30:54,240
I create the outline,

502
00:30:54,240 --> 00:30:56,320
how I want that project to appear.

503
00:30:56,320 --> 00:30:58,640
Then in VS Code,

504
00:30:58,640 --> 00:31:01,440
we have now the developer tool.

505
00:31:01,440 --> 00:31:03,760
So I use co-finals,

506
00:31:03,760 --> 00:31:06,880
they co-finals, they do talk co-finals.

507
00:31:06,880 --> 00:31:10,920
So in the data co-finals, now we have the remote.

508
00:31:10,920 --> 00:31:13,880
So we have the plan mode, we have the upload

509
00:31:13,880 --> 00:31:15,960
and we have the agent mode.

510
00:31:15,960 --> 00:31:20,960
So it's very important as a developer now,

511
00:31:20,960 --> 00:31:25,720
because honestly, I feel like these developers tools

512
00:31:25,720 --> 00:31:30,040
have been so helpful even to me as a developer,

513
00:31:30,040 --> 00:31:35,040
because now it has me now creates agents faster.

514
00:31:35,040 --> 00:31:40,280
And what I do say is being a good developer,

515
00:31:40,280 --> 00:31:43,280
you also need to be a good from team junior.

516
00:31:43,280 --> 00:31:48,280
So a good from all these things, the best outcome.

517
00:31:48,280 --> 00:31:51,480
So whenever you feel that in an idea,

518
00:31:51,480 --> 00:31:56,960
I always use the three-part team, the co-finals,

519
00:31:56,960 --> 00:31:58,360
the data co-finals.

520
00:31:58,360 --> 00:32:00,080
So where we have the ask,

521
00:32:00,080 --> 00:32:04,640
I, the ask mode, this is where I always just put in my

522
00:32:04,640 --> 00:32:06,600
random ideas, you know what?

523
00:32:06,600 --> 00:32:07,920
I did a research on this,

524
00:32:07,920 --> 00:32:10,920
this is the outline I came up with,

525
00:32:10,920 --> 00:32:15,560
this is the goal I want to achieve

526
00:32:15,560 --> 00:32:20,000
and just now brainstorm using my developer AI.

527
00:32:20,000 --> 00:32:22,560
So it always comes up with a great idea,

528
00:32:22,560 --> 00:32:25,600
because you found a little handle this flow.

529
00:32:25,600 --> 00:32:30,520
This will be the structure, now this will be the entire platform.

530
00:32:30,520 --> 00:32:35,000
So after I'm done with the ask, I should put them.

531
00:32:35,000 --> 00:32:39,640
So in plan mode, this is now where it actually does

532
00:32:39,640 --> 00:32:40,800
a complete plan.

533
00:32:40,800 --> 00:32:44,600
So after plan is created, they can be able to see

534
00:32:44,600 --> 00:32:47,400
that this is how we are going to handle the first part,

535
00:32:47,400 --> 00:32:50,920
the second pattern, until the finish.

536
00:32:50,920 --> 00:32:54,840
So then the last part is the agent part.

537
00:32:54,840 --> 00:32:59,840
So in agent, this is where now it starts the actual creation

538
00:32:59,840 --> 00:33:02,280
of these applications.

539
00:33:02,280 --> 00:33:07,040
So whenever it's back, I always take the code,

540
00:33:07,040 --> 00:33:10,760
where was it that can be that,

541
00:33:10,760 --> 00:33:14,760
now this is the human loop part.

542
00:33:14,760 --> 00:33:19,040
So whenever it wants an approval to actually do something,

543
00:33:19,040 --> 00:33:24,360
I must first verify to see that it actually what I need,

544
00:33:24,360 --> 00:33:27,960
it creates the whole ecosystem.

545
00:33:27,960 --> 00:33:32,960
So after it is done, so let's say a simple web application

546
00:33:32,960 --> 00:33:37,400
or a simple application, I always do fantastic.

547
00:33:37,400 --> 00:33:42,400
So that's local laws check that everything is actually

548
00:33:42,400 --> 00:33:44,440
working perfectly.

549
00:33:44,440 --> 00:33:48,960
So after weeks now, we go through their deployment.

550
00:33:48,960 --> 00:33:54,560
So in the pre-med, we do two types of it, so we do deploy

551
00:33:54,560 --> 00:33:59,800
our application in their, what the agent is doing,

552
00:33:59,800 --> 00:34:03,680
only we do for it in Azure.

553
00:34:03,680 --> 00:34:07,040
So after weeks now, we have a testing team,

554
00:34:07,040 --> 00:34:10,760
where we open the task group, a small group,

555
00:34:10,760 --> 00:34:15,360
to do the testing together with actually made everything

556
00:34:15,360 --> 00:34:16,680
that we want.

557
00:34:16,680 --> 00:34:21,320
And yeah, so from that, it becomes a life cycle

558
00:34:21,320 --> 00:34:23,680
going forward.

559
00:34:23,680 --> 00:34:25,680
- Awesome, awesome.

560
00:34:25,680 --> 00:34:29,400
So where did you see the biggest advantages of Azure,

561
00:34:29,400 --> 00:34:31,800
AI Foundry for enterprises?

562
00:34:31,800 --> 00:34:37,840
- So biggest thing that I noticed is,

563
00:34:40,680 --> 00:34:45,680
I feel most agents are at-campchanger.

564
00:34:45,680 --> 00:34:49,000
And it's why.

565
00:34:49,000 --> 00:34:52,880
So I'm a big follower of news,

566
00:34:52,880 --> 00:34:55,240
and this is my own personal opinion.

567
00:34:55,240 --> 00:34:57,280
I'm a big follower of news.

568
00:34:57,280 --> 00:35:02,280
And recently, now every developer uses

569
00:35:02,280 --> 00:35:06,440
these developers, either we eat, eat a copain,

570
00:35:06,440 --> 00:35:09,040
or a club, or so.

571
00:35:09,040 --> 00:35:13,400
Recently, we had a first intercofinance actually,

572
00:35:13,400 --> 00:35:18,400
you know, the oppose from the new application,

573
00:35:18,400 --> 00:35:24,840
the student application, and they're actually now in pro-class.

574
00:35:24,840 --> 00:35:27,400
So that's one thing I noted.

575
00:35:27,400 --> 00:35:31,200
The next thing I noted, okay, it came as a hype,

576
00:35:31,200 --> 00:35:36,200
and I think the public backlash actually made them

577
00:35:36,200 --> 00:35:38,120
through returning back.

578
00:35:38,120 --> 00:35:40,640
So I don't know if you had a rumor

579
00:35:40,640 --> 00:35:43,840
or anthropics removing Cloud Quads

580
00:35:43,840 --> 00:35:46,400
from its subscription.

581
00:35:46,400 --> 00:35:51,400
So it was a cashback, and due to public demand,

582
00:35:51,400 --> 00:35:55,440
they is silently giving it back,

583
00:35:55,440 --> 00:35:57,680
and why am I talking about you?

584
00:35:57,680 --> 00:36:01,800
Honestly, I can say, learning these,

585
00:36:01,800 --> 00:36:07,840
it feels like we have been rented the AI application

586
00:36:08,080 --> 00:36:11,320
and actually we don't have the full control.

587
00:36:11,320 --> 00:36:15,960
So with time, it becomes expensive.

588
00:36:15,960 --> 00:36:20,480
So how do we have this discourse

589
00:36:20,480 --> 00:36:24,880
because organization we end up paying a lot,

590
00:36:24,880 --> 00:36:29,320
and a lot of money, I can use AI application.

591
00:36:29,320 --> 00:36:34,320
So this is where I saw the hosted agent be a game changer.

592
00:36:35,880 --> 00:36:39,400
So you can be able, using hosted agents,

593
00:36:39,400 --> 00:36:43,360
and this is something that I've done just a bit with,

594
00:36:43,360 --> 00:36:47,080
is you can be able to create local models.

595
00:36:47,080 --> 00:36:49,040
So let's say for example,

596
00:36:49,040 --> 00:36:52,320
you want to run their Olam model.

597
00:36:52,320 --> 00:36:57,320
So you can be able now to run Olam locally,

598
00:36:57,320 --> 00:37:02,320
and control the license, and host it using their hosted agents.

599
00:37:04,120 --> 00:37:08,440
So for the will do the whole scaling,

600
00:37:08,440 --> 00:37:10,560
connectivity, and everything,

601
00:37:10,560 --> 00:37:14,600
and enterprises now can use it as a cheaper price.

602
00:37:14,600 --> 00:37:16,920
They only think that they will need to pay for

603
00:37:16,920 --> 00:37:20,520
the hosting price of their application.

604
00:37:20,520 --> 00:37:24,560
So this will be cheaper, and the most benefit is

605
00:37:24,560 --> 00:37:29,560
organization will have control is everything.

606
00:37:30,120 --> 00:37:35,920
So whenever people have control of how their models are run,

607
00:37:35,920 --> 00:37:39,080
how their models have been trained, I feel it.

608
00:37:39,080 --> 00:37:41,040
That's the biggest game changer,

609
00:37:41,040 --> 00:37:45,120
because from my perspective, I feel in the next couple

610
00:37:45,120 --> 00:37:48,440
of two, three years, AI is going to be expensive.

611
00:37:48,440 --> 00:37:55,400
- When you think a little bit of designing scalable architecture,

612
00:37:55,400 --> 00:37:56,600
what tips can you give?

613
00:37:57,600 --> 00:38:05,600
- So one thing I can say,

614
00:38:05,600 --> 00:38:12,600
monitoring, governance, and safety gadgets,

615
00:38:12,600 --> 00:38:19,200
because the first thing monitoring,

616
00:38:19,200 --> 00:38:24,640
monitoring now will be both in the LL monitoring,

617
00:38:24,640 --> 00:38:27,360
and that application monitoring.

618
00:38:27,360 --> 00:38:32,320
We should be able to create an application where you can be able to see

619
00:38:32,320 --> 00:38:38,480
how it's usage, how you can be able to scale up or scale down the infrastructure

620
00:38:38,480 --> 00:38:44,840
based on the traffic whenever an application gets a lot of traffic.

621
00:38:44,840 --> 00:38:49,520
How can you be able now to scale it up to accommodate this?

622
00:38:49,520 --> 00:38:55,480
How can you load that traffic?

623
00:38:55,480 --> 00:39:02,960
So that's the monitoring part of only how it is very, very beneficial.

624
00:39:02,960 --> 00:39:05,000
The next part is governance.

625
00:39:05,000 --> 00:39:08,960
And in governance, now we are talking about data security,

626
00:39:08,960 --> 00:39:13,840
data protection is by information safety,

627
00:39:13,840 --> 00:39:22,120
use of the application if a data application, or if a data application for my organization

628
00:39:22,120 --> 00:39:26,640
is any data that I have, let's say for example, in my one day, is itself

629
00:39:26,640 --> 00:39:28,800
who has access to it.

630
00:39:28,800 --> 00:39:32,400
And this is where sensitivity levels come in.

631
00:39:32,400 --> 00:39:36,080
So if you don't have access to a patch-file information,

632
00:39:36,080 --> 00:39:38,560
can you be able to access it?

633
00:39:38,560 --> 00:39:46,240
So governance brings the whole picture where it protects the user,

634
00:39:46,240 --> 00:39:48,560
the developer and the organization.

635
00:39:48,560 --> 00:39:54,760
So now that will be the data people it is, making sure that the amodes

636
00:39:54,760 --> 00:39:58,040
that you use actually are transparent and everything,

637
00:39:58,040 --> 00:40:04,640
that people are not actually using it, and that will target the next practice.

638
00:40:04,640 --> 00:40:09,440
So sitting up from practice, that will help you

639
00:40:09,440 --> 00:40:14,880
help your application to use a bone file.

640
00:40:14,880 --> 00:40:15,680
Awesome.

641
00:40:15,680 --> 00:40:21,360
Can you a little bit explain how a Azure AI Foundry integrates,

642
00:40:21,360 --> 00:40:27,920
like logic apps, Q1nets, power platform, and so on?

643
00:40:27,920 --> 00:40:37,840
So how I have done this, hungry users, MCP client protocol.

644
00:40:37,840 --> 00:40:44,240
So these client protocol, it can be able now to connect with anything.

645
00:40:44,240 --> 00:40:48,640
So let's say for example, you want to connect to power platform.

646
00:40:48,640 --> 00:40:52,120
So hungry has the MCP connector, power platform,

647
00:40:52,120 --> 00:40:56,640
you can be able to connect directly to logic apps.

648
00:40:56,640 --> 00:41:00,640
So you can either connect it to power automate.

649
00:41:00,640 --> 00:41:09,200
So we have the automate in a Foundry, you can either use that flow automation

650
00:41:09,200 --> 00:41:16,360
that is in Foundry, or you can also call the logic app using the MCP next.

651
00:41:16,360 --> 00:41:19,520
You also talk about content organization.

652
00:41:19,520 --> 00:41:26,080
So that's how Foundry connects with all these external applications.

653
00:41:26,080 --> 00:41:38,400
When I think so a little bit outside from the Microsoft tools and so on, services,

654
00:41:38,400 --> 00:41:44,440
there are, can you, yeah, I don't know, like sales, also,

655
00:41:44,440 --> 00:41:47,160
on some other enterprise solutions.

656
00:41:47,160 --> 00:41:52,720
Can you connect them and can make it security?

657
00:41:52,720 --> 00:41:59,040
Yeah, actually you can be able to connect with the support from Foundry because

658
00:41:59,040 --> 00:42:06,320
I honestly, I'll say, MCP is changing how the edges connect because now you have that

659
00:42:06,320 --> 00:42:14,160
trutting layer, you can be able to call in any application that is outside Microsoft ecosystem.

660
00:42:14,160 --> 00:42:18,720
So MCP has been the biggest in changing this.

661
00:42:18,720 --> 00:42:22,160
Yeah, that's what I can say about it.

662
00:42:22,160 --> 00:42:30,480
And I see what also interesting topic is I see the retrieval augmented generation of the

663
00:42:30,480 --> 00:42:32,080
big topic actually.

664
00:42:32,080 --> 00:42:33,080
Can you agree?

665
00:42:33,080 --> 00:42:36,160
Can you talk about the architecture in Azure?

666
00:42:36,160 --> 00:42:40,360
So the architecture of RAC?

667
00:42:40,360 --> 00:42:42,960
Oh, the architecture of RAC.

668
00:42:42,960 --> 00:42:44,960
So I think so.

669
00:42:44,960 --> 00:42:51,440
So RAC, RAC is there, retrieval of minted generation.

670
00:42:51,440 --> 00:42:58,880
This is where you can be able now to actually expound on the knowledge of how to explain

671
00:42:58,880 --> 00:43:00,880
what these are.

672
00:43:00,880 --> 00:43:03,520
Let imagine this.

673
00:43:03,520 --> 00:43:13,440
So let's say, for example, you left work yesterday at 5 pm and when you left work, let's

674
00:43:13,440 --> 00:43:25,760
say at that particular time, let me explain this in.

675
00:43:25,760 --> 00:43:27,760
Okay, how can I do this?

676
00:43:27,760 --> 00:43:33,360
So I'm looking for the best explanation for this.

677
00:43:33,360 --> 00:43:39,600
So yeah, this is the best explanation of RAC.

678
00:43:39,600 --> 00:43:42,880
So it's my explanation.

679
00:43:42,880 --> 00:43:45,920
So yeah, when were you in school?

680
00:43:45,920 --> 00:43:52,480
We were told that, you know what the planet is, this whole system has length.

681
00:43:52,480 --> 00:43:59,200
So, you know, the material mass, plan, the length planet.

682
00:43:59,200 --> 00:44:06,160
So bring up and you'll just know that, you know what, this whole system has length planet.

683
00:44:06,160 --> 00:44:11,440
When a new kid comes and asks if got how many planets are in this whole system, you see

684
00:44:11,440 --> 00:44:12,440
that?

685
00:44:12,440 --> 00:44:13,440
You know what?

686
00:44:13,440 --> 00:44:20,280
From main knowledge, I know that this whole system has length planets, but from the current

687
00:44:20,280 --> 00:44:23,920
data, it goes in that loop is for longer planets.

688
00:44:23,920 --> 00:44:28,280
So that current information that I don't have.

689
00:44:28,280 --> 00:44:37,000
So the same with large language models, where the large language models are trained to

690
00:44:37,000 --> 00:44:43,720
have a particular date.

691
00:44:43,720 --> 00:44:49,000
So let's say for example, the GBT models, they are trained up to, let's say, 2024 or 2025.

692
00:44:49,000 --> 00:44:56,600
So whenever you ask a question that is here in 2020, it cannot be able to find it.

693
00:44:56,600 --> 00:44:59,160
So this is where RAC comes in.

694
00:44:59,160 --> 00:45:08,280
RAC comes in with the new information and it updates the LLM data strength.

695
00:45:08,280 --> 00:45:16,840
So the LLM was last trained on, let's say, December of 2025 and this new thing that has

696
00:45:16,840 --> 00:45:17,840
happened.

697
00:45:17,840 --> 00:45:21,800
So it will go ahead and tell if you know what, this is the new.

698
00:45:21,800 --> 00:45:28,080
So in my case, a new, there are length planets and this, it comes and tells me, you know

699
00:45:28,080 --> 00:45:32,560
what from our syllabus, we hear that the previous data planet.

700
00:45:32,560 --> 00:45:37,680
So now I update my knowledge and be like, you know what, the latest data planet.

701
00:45:37,680 --> 00:45:42,720
So in now, organizational, how does RAC work?

702
00:45:42,720 --> 00:45:50,680
So RAC in an organization, this is where we are able now to connect with and enterprise

703
00:45:50,680 --> 00:45:51,680
knowledge.

704
00:45:51,680 --> 00:45:57,800
So we can be able, we have a lot of data in our workplace.

705
00:45:57,800 --> 00:46:02,880
So and these data are in style also, they are separated.

706
00:46:02,880 --> 00:46:07,800
So for me to find a particular information, Italy always take an out of them.

707
00:46:07,800 --> 00:46:13,560
So myself, for example, I have a document that is in shabot, another document is in one

708
00:46:13,560 --> 00:46:19,440
drive, you know, another document, but attached in my email.

709
00:46:19,440 --> 00:46:26,960
So for me to be able to address all this information, we can be able to connect RAC.

710
00:46:26,960 --> 00:46:38,320
So this is now when I can be able now to now call the shabot data bring in those information,

711
00:46:38,320 --> 00:46:41,600
bring the one data, bring the output data.

712
00:46:41,600 --> 00:46:45,520
Or whenever I have stored my information.

713
00:46:45,520 --> 00:46:52,600
So I will go ahead and ask what is the address code in my company.

714
00:46:52,600 --> 00:46:59,080
And with that information, it can be able to understand the context because now, let's

715
00:46:59,080 --> 00:47:09,880
say, for example, I go to my fairbite, I model and ask what does Edgar like or does where

716
00:47:09,880 --> 00:47:13,680
every day it doesn't have that information about the guy.

717
00:47:13,680 --> 00:47:19,880
So for it to know the information about me, I'll have to train it.

718
00:47:19,880 --> 00:47:27,120
So give me information, you know what this is, there is a theme of Edgar, this is how Edgar

719
00:47:27,120 --> 00:47:32,040
writes his book, this is how Edgar's writing his article.

720
00:47:32,040 --> 00:47:45,160
So now me giving this language knowledge my data, it can now be able to understand that

721
00:47:45,160 --> 00:47:50,760
the context, so whenever I ask what Edgar's favorite food can be able to answer that.

722
00:47:50,760 --> 00:47:57,120
So send to an organization, whenever we give these organization information and an

723
00:47:57,120 --> 00:48:01,080
influence actually, they can actually are querying it.

724
00:48:01,080 --> 00:48:06,720
So it brings us that rich content.

725
00:48:06,720 --> 00:48:15,080
And when you have this big project, you have also select vector data basis, search solution,

726
00:48:15,080 --> 00:48:29,000
how is your model to choose the right solutions because there are so many included in error.

727
00:48:29,000 --> 00:48:33,120
So now that's the indexing plan.

728
00:48:33,120 --> 00:48:40,640
So whenever I now let's say we have data in Shackold and we want our location to retrieve

729
00:48:40,640 --> 00:48:41,640
this data.

730
00:48:41,640 --> 00:48:46,280
So the first thing is always indexing this data and there are two ways you can be able to

731
00:48:46,280 --> 00:48:47,760
index the data.

732
00:48:47,760 --> 00:48:51,120
So the first one is a semantic index.

733
00:48:51,120 --> 00:48:59,600
So a semantic index is where you have what that has correlated, you know, one one can have

734
00:48:59,600 --> 00:49:00,600
a lot of meaning.

735
00:49:00,600 --> 00:49:02,600
So it does that.

736
00:49:02,600 --> 00:49:04,600
It actually is free.

737
00:49:04,600 --> 00:49:05,600
We have two ways.

738
00:49:05,600 --> 00:49:07,200
We should can index that.

739
00:49:07,200 --> 00:49:09,200
So the first one is P1.

740
00:49:09,200 --> 00:49:14,800
So in P1, this is where it touches for a particular keyword.

741
00:49:14,800 --> 00:49:23,800
So whenever I touch for a net set is that it just goes to that SharePoint connector for

742
00:49:23,800 --> 00:49:28,920
the index that you created and touches for the keyword pizza.

743
00:49:28,920 --> 00:49:30,640
And next is the semantic.

744
00:49:30,640 --> 00:49:36,800
So in semantic is a combination for the keyword and the semantic.

745
00:49:36,800 --> 00:49:42,280
So it can be able to and that's not the best but yeah, it does.

746
00:49:42,280 --> 00:49:47,480
So the best way to think is that's how do it is a factor such.

747
00:49:47,480 --> 00:49:51,800
So we've always created and embedding using the translation.

748
00:49:51,800 --> 00:49:59,080
So in the translation from this is where now we divide this data into Shackold.

749
00:49:59,080 --> 00:50:04,760
So once in zero and it now the A A can be able to understand it.

750
00:50:04,760 --> 00:50:12,680
So whenever you have a question, it goes to that factor index and the best response for

751
00:50:12,680 --> 00:50:13,680
you.

752
00:50:13,680 --> 00:50:20,680
So how often is the best with always use.

753
00:50:20,680 --> 00:50:23,680
Combination of vector and semantic.

754
00:50:23,680 --> 00:50:25,480
It always for specific.

755
00:50:25,480 --> 00:50:27,480
Yes, you can.

756
00:50:27,480 --> 00:50:28,480
Awesome.

757
00:50:28,480 --> 00:50:33,400
And I think there when I think about enterprise, there are a lot of.

758
00:50:33,400 --> 00:50:34,720
Yeah, there are different countries.

759
00:50:34,720 --> 00:50:37,080
Is there any.

760
00:50:37,080 --> 00:50:43,480
Option to do a multi region AI architecture in.

761
00:50:43,480 --> 00:50:44,480
Foundry.

762
00:50:44,480 --> 00:50:48,240
Yeah, actually, yeah, about that.

763
00:50:48,240 --> 00:50:51,040
So.

764
00:50:51,040 --> 00:50:59,720
The first thing that we created using mella or interlonal application was mella was tenant

765
00:50:59,720 --> 00:51:04,720
model.

766
00:51:04,720 --> 00:51:08,240
Okay, with the simplest, but it's.

767
00:51:08,240 --> 00:51:09,720
At least simple.

768
00:51:09,720 --> 00:51:13,080
So because everything is authenticated.

769
00:51:13,080 --> 00:51:17,440
And yeah, so now the real challenge.

770
00:51:17,440 --> 00:51:21,640
So single channel, I can say is actually isolated.

771
00:51:21,640 --> 00:51:25,080
So it's the logic is simple.

772
00:51:25,080 --> 00:51:27,680
How it created that type is.

773
00:51:27,680 --> 00:51:34,160
So the biggest biggest challenge has been now creating an interface.

774
00:51:34,160 --> 00:51:36,440
So because interface.

775
00:51:36,440 --> 00:51:40,680
You need to really care about security.

776
00:51:40,680 --> 00:51:44,440
How are people or how are the organization.

777
00:51:44,440 --> 00:51:47,640
Create or use these data.

778
00:51:47,640 --> 00:51:49,440
Are there any leakages?

779
00:51:49,440 --> 00:51:52,640
So we talk about what I saw mentioned.

780
00:51:52,640 --> 00:51:55,080
We talk about people.

781
00:51:55,080 --> 00:52:00,080
And make sure that the partitions are.

782
00:52:00,080 --> 00:52:07,320
So what we did for mella, I are creating the enterprise there.

783
00:52:07,320 --> 00:52:10,320
We actually had it now do.

784
00:52:10,320 --> 00:52:13,840
They did that you see.

785
00:52:13,840 --> 00:52:16,320
Isolation well each claim.

786
00:52:16,320 --> 00:52:17,960
We carried the.

787
00:52:17,960 --> 00:52:32,960
And let's take the we carried make every information about this point.

788
00:52:32,960 --> 00:52:38,960
So this is where it can only get that that is a soul.

789
00:52:38,960 --> 00:52:45,960
Access to so the next one is partitioning make making sure that each tenant has a different

790
00:52:45,960 --> 00:52:51,280
position and they cannot access any information that is in a different tenant.

791
00:52:51,280 --> 00:52:58,960
So building an interface application is challenging, but you really really need to ensure

792
00:52:58,960 --> 00:53:02,200
that you have good governance.

793
00:53:02,200 --> 00:53:04,360
You need you have to decide.

794
00:53:04,360 --> 00:53:14,280
And when you think and continue still continuous integration continues delivery strategy.

795
00:53:14,280 --> 00:53:21,280
How does it work for AI applications, especially.

796
00:53:21,280 --> 00:53:26,280
So we do it to use the GitHub action.

797
00:53:26,280 --> 00:53:32,280
So we use it up for the continuous implementation and continuous development.

798
00:53:32,280 --> 00:53:40,280
So whenever I knew application or we have features that we really need to find.

799
00:53:40,280 --> 00:53:44,280
So first thing we do that testing.

800
00:53:44,280 --> 00:53:48,280
So we have the two more so we have the production and the testing.

801
00:53:48,280 --> 00:53:54,280
So in testing we always do testing, making sure that everything works well.

802
00:53:54,280 --> 00:54:00,280
So everything is given go ahead.

803
00:54:00,280 --> 00:54:07,280
So the continuous integration continuous line is this where now we always update the current

804
00:54:07,280 --> 00:54:11,280
and we take care of that.

805
00:54:11,280 --> 00:54:19,280
We have talked about a lot of work a lot in AI.

806
00:54:19,280 --> 00:54:30,280
Any good option to monitor it, I think especially on when I talk to the CFO.

807
00:54:30,280 --> 00:54:40,280
I think the C is the cost of the adding I think financial financial ops or something I can monitor.

808
00:54:40,280 --> 00:54:44,280
Yeah, the workload and the cost.

809
00:54:44,280 --> 00:54:46,280
You can actually do that.

810
00:54:46,280 --> 00:54:49,280
And that's one of the thing we do.

811
00:54:49,280 --> 00:55:03,280
So we sell the Mela as an interface of past.

812
00:55:03,280 --> 00:55:09,280
We can get the have Mela be we sell it as a yes, that's for the service or.

813
00:55:09,280 --> 00:55:13,280
We can deploy it to our clients.

814
00:55:13,280 --> 00:55:19,280
We have the hosting for the models, but it's quite cheaper.

815
00:55:19,280 --> 00:55:31,280
So as hosting it always provides cheaper chance for the clients because now they will not worry about having to check every time.

816
00:55:31,280 --> 00:55:39,280
When anything fails, we always are on top of it and always checking up always updating it.

817
00:55:39,280 --> 00:55:53,280
So monitoring has been it's very easy to do because you can create a good application and yeah, you can get a lot of money for it.

818
00:55:53,280 --> 00:55:54,280
Okay, cool.

819
00:55:54,280 --> 00:55:57,280
So oh, we are running out of time.

820
00:55:57,280 --> 00:55:58,280
I hope.

821
00:55:58,280 --> 00:55:59,280
Yeah, I have.

822
00:55:59,280 --> 00:56:03,280
I go to the hot stake questions.

823
00:56:03,280 --> 00:56:13,280
Most companies don't need fine cute models. They need better data grounding to your agree disagree and why.

824
00:56:13,280 --> 00:56:18,280
I've been working on tuning and data grounding by a doctor for data grounding.

825
00:56:18,280 --> 00:56:20,280
So yeah, agree.

826
00:56:20,280 --> 00:56:25,280
Because in data grounding, this is where I have control.

827
00:56:25,280 --> 00:56:35,280
So I have control or what I want to make a show to do. I have control of data that I keep it and you want to know.

828
00:56:35,280 --> 00:56:39,280
So yeah, I do absolutely agree.

829
00:56:39,280 --> 00:56:41,280
Yeah, last I heard.

830
00:56:41,280 --> 00:56:46,280
I governance model models will become more important than model quality.

831
00:56:46,280 --> 00:56:51,280
What impact will this have on enterprise AI strategies.

832
00:56:51,280 --> 00:56:54,280
Model governance.

833
00:56:54,280 --> 00:56:58,280
This is always the best because for model governance.

834
00:56:58,280 --> 00:57:02,280
We need to protect that they use that.

835
00:57:02,280 --> 00:57:05,280
Either they use that they don't want that conversation.

836
00:57:05,280 --> 00:57:09,280
So for the user.

837
00:57:09,280 --> 00:57:12,280
Does he have.

838
00:57:12,280 --> 00:57:24,280
And the conditions to access that.

839
00:57:24,280 --> 00:57:28,280
Can you access information that he or she or they don't have.

840
00:57:28,280 --> 00:57:33,280
The permission to do that. So in governance, you can see the usage of the eye.

841
00:57:33,280 --> 00:57:37,280
I will be thinking use to is actually was using this AI.

842
00:57:37,280 --> 00:57:43,280
Is it really.

843
00:57:43,280 --> 00:57:47,280
Is it doing.

844
00:57:47,280 --> 00:57:51,280
And we always go for.

845
00:57:51,280 --> 00:58:01,280
Okay, small language models will replace many expensive GTP work loads, which enterprise use cases are the best for.

846
00:58:01,280 --> 00:58:02,280
We are.

847
00:58:02,280 --> 00:58:07,280
Much best for.

848
00:58:07,280 --> 00:58:08,280
This.

849
00:58:08,280 --> 00:58:10,280
This.

850
00:58:10,280 --> 00:58:12,280
And for me, yeah.

851
00:58:12,280 --> 00:58:14,280
And I've said it.

852
00:58:14,280 --> 00:58:17,280
More language models have different than because.

853
00:58:17,280 --> 00:58:20,280
Less language models for the close motors.

854
00:58:20,280 --> 00:58:23,280
Will be expensive and it feels like.

855
00:58:23,280 --> 00:58:26,280
We are renting. And you know where you're renting something.

856
00:58:26,280 --> 00:58:31,280
You don't have control because somebody some day will just.

857
00:58:31,280 --> 00:58:34,280
And I say, you know what you don't have access to the.

858
00:58:34,280 --> 00:58:36,280
You cannot access this model.

859
00:58:36,280 --> 00:58:39,280
And because I want access and control of my modern.

860
00:58:39,280 --> 00:58:41,280
It limited course.

861
00:58:41,280 --> 00:58:43,280
And with.

862
00:58:43,280 --> 00:58:45,280
My own customization.

863
00:58:45,280 --> 00:58:47,280
And always, always.

864
00:58:47,280 --> 00:58:48,280
For.

865
00:58:48,280 --> 00:58:49,280
Small.

866
00:58:49,280 --> 00:58:51,280
Small language models.

867
00:58:51,280 --> 00:58:52,280
Okay.

868
00:58:52,280 --> 00:58:57,280
And from engineering will eventually become less important them or cast ring and.

869
00:58:57,280 --> 00:58:59,280
s workplow design, do you think

870
00:58:59,280 --> 00:59:01,900
this shift is already happening here?

871
00:59:01,900 --> 00:59:05,120
>> Yeah, I think it is very dead,

872
00:59:05,120 --> 00:59:10,440
because how I can describe from

873
00:59:10,440 --> 00:59:13,960
engineering is currently, if we

874
00:59:13,960 --> 00:59:17,040
want, I can say 80% of given away I'm

875
00:59:17,040 --> 00:59:19,320
using iTunes.

876
00:59:19,320 --> 00:59:23,180
So, this big thread,

877
00:59:23,180 --> 00:59:26,440
in my own explanation, I can say,

878
00:59:26,440 --> 00:59:30,440
being a good-of-one engineer is being a good

879
00:59:30,440 --> 00:59:32,640
different offer, that's my own take,

880
00:59:32,640 --> 00:59:37,480
because a good product will always give a good product.

881
00:59:37,480 --> 00:59:42,760
So, when we are using the application for AI

882
00:59:42,760 --> 00:59:46,960
coding agents, and we give it a good prompt,

883
00:59:46,960 --> 00:59:49,840
it always helps for tokens, it always helps

884
00:59:49,840 --> 00:59:55,240
from that content, and we always get a good project,

885
00:59:55,240 --> 00:59:57,000
or a good response.

886
00:59:57,000 --> 01:00:03,800
So, a good prompt in engineering will bring a good design.

887
01:00:03,800 --> 01:00:07,240
>> Awesome. Now, my final closing,

888
01:00:07,240 --> 01:00:09,000
if you're leading an enterprise,

889
01:00:09,000 --> 01:00:11,840
AI transformation project using AI Foundry,

890
01:00:11,840 --> 01:00:14,960
what will be your first three

891
01:00:14,960 --> 01:00:17,880
prioratories and why?

892
01:00:17,880 --> 01:00:20,840
>> Monitoring.

893
01:00:20,840 --> 01:00:23,880
Monitoring because now,

894
01:00:23,880 --> 01:00:26,080
my priority will be monitoring,

895
01:00:26,080 --> 01:00:29,760
I want to know the users who are actually using

896
01:00:29,760 --> 01:00:36,280
this application, the frequency they are using it,

897
01:00:36,280 --> 01:00:40,560
I want to monitor the usage,

898
01:00:40,560 --> 01:00:43,840
because it will protect both me and them,

899
01:00:43,840 --> 01:00:46,840
for me, it will protect both course,

900
01:00:46,840 --> 01:00:51,840
and I can be able to quickly know whenever something is

901
01:00:51,840 --> 01:00:54,880
not actually the one thing I want.

902
01:00:54,880 --> 01:00:59,920
So, I honestly feel like monitoring

903
01:00:59,920 --> 01:01:03,440
and the interface of the application is the biggest part.

904
01:01:03,440 --> 01:01:08,960
>> Awesome. So, yeah, thank you so much for your time.

905
01:01:08,960 --> 01:01:12,120
I hope we can do another session in deep dive

906
01:01:12,120 --> 01:01:17,720
more in security governance and the responsibility AI sometime.

907
01:01:17,720 --> 01:01:19,880
Thank you so much.

908
01:01:19,880 --> 01:01:23,560
This was a really interesting deep dive. Thank you.

909
01:01:23,560 --> 01:01:26,200
>> Yeah, thank you so much. It's my pleasure.

910
01:01:26,200 --> 01:01:28,800
And I'm happy to be here.

Mirko Peters

Founder of m365.fm, m365.show and m365con.net

Mirko Peters is a Microsoft 365 expert, content creator, and founder of m365.fm, a platform dedicated to sharing practical insights on modern workplace technologies. His work focuses on Microsoft 365 governance, security, collaboration, and real-world implementation strategies.

Through his podcast and written content, Mirko provides hands-on guidance for IT professionals, architects, and business leaders navigating the complexities of Microsoft 365. He is known for translating complex topics into clear, actionable advice, often highlighting common mistakes and overlooked risks in real-world environments.

With a strong emphasis on community contribution and knowledge sharing, Mirko is actively building a platform that connects experts, shares experiences, and helps organizations get the most out of their Microsoft 365 investments.

Edgar McOchieng

I'm a Data & AI Engineer and Microsoft MVP (Microsoft Foundry and Business Applications) based in Nairobi, Kenya. I build enterprise AI products on Azure and run McOchieng Labs, a free developer community focused on mentoring university students into the industry. I'm passionate about making sure developers across Africa have real access to the tools shaping the industry

Building and deploying production grade AI agents with Microsoft Foundry with Edgar McOchieng [MVP]