From Data to Intelligent Agents: Building Trusted Enterprise AI with Microsoft AI Foundry with Shubhangi Goyal [MVP]

1
00:00:00,000 --> 00:00:05,600
Hello everyone, welcome back to another episode of VM 365 FAM podcast.

2
00:00:05,600 --> 00:00:08,720
I'm your host, Mocropitas, and today we are diving into one of the

3
00:00:08,720 --> 00:00:12,720
hottest topics in the Microsoft ecosystem, Enterprise AI.

4
00:00:12,720 --> 00:00:17,120
Over the last two years, we all have been talking about

5
00:00:17,120 --> 00:00:22,080
ChatGPT, Co-Pilot, and Generative AI, but building AI that

6
00:00:22,080 --> 00:00:26,080
actually creates business value in certain organizations, we cry out much more

7
00:00:26,080 --> 00:00:27,840
than great prompts.

8
00:00:27,840 --> 00:00:31,280
Today is guest understand exactly what happened behind the scene.

9
00:00:31,280 --> 00:00:35,920
Joining me is Shupag, Yiyal, Microsoft MVP, International,

10
00:00:35,920 --> 00:00:39,280
Rigenized Speaker, Zenyur, Daychallant, AI professional, and

11
00:00:39,280 --> 00:00:42,880
somewhat, we have spent more than indicate helping organizations

12
00:00:42,880 --> 00:00:46,080
turn data into intelligent decision making.

13
00:00:46,080 --> 00:00:49,920
We'll explore Microsoft AI Foundry, Azure, OpenAI,

14
00:00:49,920 --> 00:00:52,640
Data Science, Responsible AI, Enterprise Adoptions,

15
00:00:52,640 --> 00:00:56,720
the future of intelligent agents, and why good AI

16
00:00:56,720 --> 00:00:59,040
always start with good data.

17
00:00:59,040 --> 00:01:04,080
Yeah, Shupag, you're welcome to the NC65s as N, it's fantastic to have you here.

18
00:01:04,080 --> 00:01:09,040
Thank you, great to be here, and I'm looking forward to sharing all the insights

19
00:01:09,040 --> 00:01:12,560
that I have in this area. And yeah, of course, AI has really,

20
00:01:12,560 --> 00:01:17,360
I would say, full of load now in the past few years, and

21
00:01:17,360 --> 00:01:22,320
we are seeing literally being a part of everything we could potentially imagine.

22
00:01:22,320 --> 00:01:27,920
Yeah, yeah, first thing I congratulate you to your RENU and the MVP award.

23
00:01:27,920 --> 00:01:38,320
Can you tell us a little bit about you, how you get into the data world,

24
00:01:38,320 --> 00:01:44,960
data science world, and especially how you change to AI, and why Microsoft?

25
00:01:44,960 --> 00:01:48,000
Yeah, actually, that's quite interesting.

26
00:01:48,000 --> 00:01:53,760
So I think around 10 years ago, I would say at least more like 13 actually.

27
00:01:53,760 --> 00:01:57,760
So I pursued computer science engineering, and there was my first degree,

28
00:01:57,760 --> 00:02:03,680
and when I graduated from my course, I, at the time,

29
00:02:03,680 --> 00:02:07,600
like for example, data science was like one of the most, I would say,

30
00:02:07,600 --> 00:02:12,800
booming career arena. And when I joined my first job, which was back in India,

31
00:02:12,800 --> 00:02:16,240
I basically had the opportunity to work on some of the projects,

32
00:02:16,240 --> 00:02:20,560
which really sparked my interest in that area. And it led me to basically work on some of

33
00:02:20,560 --> 00:02:24,640
the projects there, which were mostly around like, you know, building machine learning models,

34
00:02:24,640 --> 00:02:28,240
and like, you know, analyzing data and seeing how predictive modeling works.

35
00:02:28,240 --> 00:02:31,600
So that's how I went into the world of, you know, data.

36
00:02:31,600 --> 00:02:34,720
And it became something which is like, I really enjoy working with.

37
00:02:34,720 --> 00:02:40,000
I can see like how, if you use effectively, that as a part of your, like, you know,

38
00:02:40,000 --> 00:02:44,320
as a business, how it can create impact, and how can find opportunities.

39
00:02:44,320 --> 00:02:49,520
And then like about seven years ago, I moved to the UK, and I graduated from University of

40
00:02:49,520 --> 00:02:53,520
Pat, I pursued my master's there. So I sort of moved from, I would say,

41
00:02:53,520 --> 00:02:58,960
a share to Europe in a way. And then my first job, I worked for an AI startup.

42
00:02:58,960 --> 00:03:05,040
And they are basically from data science, I entered into the arena of AI, and the rest is,

43
00:03:05,040 --> 00:03:10,960
as I said, like history. And I think one of the interesting parts of working in that job was like,

44
00:03:10,960 --> 00:03:16,080
I learned a lot more about AI, what AI really means. And I think when I joined that role,

45
00:03:16,080 --> 00:03:21,360
I thought that, you know, this is something just going to happen in maybe some years of time,

46
00:03:21,360 --> 00:03:25,360
but actually happened a lot before. And I was working on some of the projects there,

47
00:03:25,360 --> 00:03:30,640
which were already in the space, like for example, for example, wise agents, or like,

48
00:03:30,640 --> 00:03:36,320
building web char agents and so on. And then when AI came, it actually expanded that a lot more.

49
00:03:36,320 --> 00:03:42,800
And I just ended up like where I am. And yeah, like I have basically worked a lot of Microsoft tech

50
00:03:42,800 --> 00:03:48,880
in the Microsoft ecosystem. And also with Microsoft as like the customer in my first job. So it's like

51
00:03:48,880 --> 00:03:53,920
a lot of, I will say like using the ecosystem entirely, you know, working on Azure, for example,

52
00:03:53,920 --> 00:03:59,600
or even using data platform tooling and also like working now on Foundry. So yeah,

53
00:03:59,600 --> 00:04:03,760
this is how basically my journey began. I think it was more like, you know, it's just,

54
00:04:03,760 --> 00:04:08,240
I wouldn't say it's like I venture into AI. It's just like, you know, happened a longer way. And like,

55
00:04:08,240 --> 00:04:13,360
I enjoyed it so much that I just thought, you know, this is like something I would want to be a part

56
00:04:13,360 --> 00:04:17,600
of and remain a part of for as long as I want to be. So yeah, this is how it ended up.

57
00:04:17,600 --> 00:04:27,600
Yeah, yeah, yeah. You have said you have started at the start up. And now you are in the enterprise world.

58
00:04:28,640 --> 00:04:35,360
How different is working in these both worlds? So obviously like if we talk about working like

59
00:04:35,360 --> 00:04:40,560
big corporate enterprise world versus working in startups, the working cell is quite different.

60
00:04:40,560 --> 00:04:44,560
Startups are definitely higher-paced and like, you know, faster-paced, no matter where you work. And

61
00:04:44,560 --> 00:04:50,480
enterprise are like most sort of like the move, but like you have to be like, you know, they have to

62
00:04:50,480 --> 00:04:55,760
be like super like careful with the fact that you know, when we are moving, making changes,

63
00:04:55,760 --> 00:04:59,840
how does that impact like the entire organization? And then like when you work on start this,

64
00:04:59,840 --> 00:05:04,960
also you can have like access to a lot more, I would say cutting edge technologies, which I still

65
00:05:04,960 --> 00:05:08,880
have, but I think startups because they can move faster, they can experiment a lot faster than

66
00:05:08,880 --> 00:05:13,520
that's what I can say at least. In like big corporates, like you will have to make sure that,

67
00:05:13,520 --> 00:05:18,800
for example, even if you're performing a migration, then like, how does that migration impact the

68
00:05:18,800 --> 00:05:24,400
entire, for example, structure of an approach that you have built? How, how to change? And then

69
00:05:24,400 --> 00:05:29,120
what's the ultimate goal we are trying to check? So I would say enterprise AI has, you know,

70
00:05:29,120 --> 00:05:34,640
where if I see it like, evolved a lot more in the sense like you can see that previously people

71
00:05:34,640 --> 00:05:38,960
were experimenting with it, but now it has reached a stage of maturity with organizations,

72
00:05:38,960 --> 00:05:44,880
actively want to derive value from it. And having said that I know that there has been a lot of,

73
00:05:44,880 --> 00:05:49,120
you know, discussion going on that, you know, a lot of people don't see benefits and so on.

74
00:05:49,120 --> 00:05:53,840
But I think it starts with recognizing the right use case of where you can implement AI.

75
00:05:53,840 --> 00:05:59,840
Like are you trying to aim that, okay, I have this product I'm going to create or build with AI,

76
00:05:59,840 --> 00:06:06,160
but I'm expecting, for example, this ROI on this or I'm expecting that, for example, the effort of

77
00:06:06,160 --> 00:06:12,640
the process I'm currently using, it can be optimized or can be made better. So thinking about enterprise

78
00:06:12,640 --> 00:06:17,520
AI, it starts with that mindset that, okay, this is the use case and also be slightly realistic about

79
00:06:17,520 --> 00:06:24,720
it that at the end of the day, like, do we have in a free sources time and, you know, right people and,

80
00:06:24,720 --> 00:06:29,600
you know, the right platform to build it because there are robust designs really important when you

81
00:06:29,600 --> 00:06:34,640
build the system. It's not just about having a large language panel. It's also about having a

82
00:06:34,640 --> 00:06:40,800
robust design that actually, you know, caters to the need of like, whatever your goal is with that AI

83
00:06:40,800 --> 00:06:47,200
application that you are trying to build. Yeah, I say every company or everyone,

84
00:06:47,200 --> 00:06:57,680
everyone wants AI today, but how important is the data quality or was stayed with the data before I

85
00:06:57,680 --> 00:07:06,720
buy tools or edit? So data quality is a really, very big part of it. So today you design, like,

86
00:07:06,720 --> 00:07:10,800
for example, you build an AI agent or you go in the world of agent together, right, it still operates

87
00:07:10,800 --> 00:07:15,600
on some sort of data, enterprise data, customer data, whatever your organizational data is. It's not

88
00:07:15,600 --> 00:07:21,440
happening just that you can take some lush language model, plug it into your application, build like a

89
00:07:21,440 --> 00:07:26,720
front end and so on. It wouldn't work that way. And a lot of times, like, the data is not structured or

90
00:07:26,720 --> 00:07:32,000
it is not like the quality is not taken care of in a way that AI needs because, you know, it has

91
00:07:32,000 --> 00:07:37,840
hallucinates and there are times you might not get relevant outcomes or the right processes.

92
00:07:37,840 --> 00:07:42,320
And so it's super important to make sure that which people right now call this sort of context

93
00:07:42,320 --> 00:07:47,120
engineering, like, that's I think I would say it's a little bit like overlap there. So you have to

94
00:07:47,120 --> 00:07:53,600
make sure that the right context is provided to your AI application in the first place to make it

95
00:07:53,600 --> 00:07:58,320
like useful and also reliable for things. So today you are building an application that serves

96
00:07:58,320 --> 00:08:03,840
different arena of customers, maybe like there are students, they could be like people who are in

97
00:08:03,840 --> 00:08:08,880
mid-age or maybe like older population that you have to make sure that it is able to cater to

98
00:08:08,880 --> 00:08:14,880
this wide set of audience and, you know, is actually user friendly. I would say like user experience is

99
00:08:14,880 --> 00:08:19,840
great and then also like, big and trust it in the first place and that's where responsibility

100
00:08:19,840 --> 00:08:25,840
I ventures into the picture. And for most part, if you are a B2B company, right, like, the other

101
00:08:25,840 --> 00:08:30,480
business is going to purchase your product because they want to also understand how you have built

102
00:08:30,480 --> 00:08:35,760
the product firstly, then I think they would also want to know like how your AI has entered

103
00:08:35,760 --> 00:08:40,240
up like making some decisions, which I think is called sort of like an expandable AI like.

104
00:08:40,240 --> 00:08:45,520
So if you can't explain how it has worked, I don't think they would be able to trust it entirely. So

105
00:08:45,520 --> 00:08:52,160
it's super important these small aspects while building it. And then data quality is another key area.

106
00:08:52,160 --> 00:08:57,760
So for example, if I'm asking similar set of questions every single time or testing it, right,

107
00:08:57,760 --> 00:09:02,400
would it give me similar outcomes? Will it change? And then how much would that change be? And

108
00:09:02,400 --> 00:09:09,600
me as a customer, am I happy to have that change? I think those are some nuances which you can only

109
00:09:09,600 --> 00:09:14,560
know when you test it enough, but even like after building it. So I would say there is like a whole

110
00:09:14,560 --> 00:09:18,240
lot of process involved to ensure data quality is at the right point.

111
00:09:18,240 --> 00:09:28,160
And when we have data quality, what other things a company needs to be a, or a chief AI readiness?

112
00:09:29,840 --> 00:09:34,960
So AI readiness is a very broad term and there's so many things that fall under AI readiness.

113
00:09:34,960 --> 00:09:40,560
Like see, there are things for example, one side is like you are building things because you want to

114
00:09:40,560 --> 00:09:46,560
like make sure that your company is like, or organization is like using AI effectively.

115
00:09:46,560 --> 00:09:51,520
Then there is a second side where you also want to use it, which is like people call AI adoption sort

116
00:09:51,520 --> 00:09:56,800
of. And everyone is up to speed, they know for example, whatever the data stool is in market.

117
00:09:56,800 --> 00:10:01,120
So let's say if I'm using co-pilot, you should know how to use co-pilot for example, or actually

118
00:10:01,120 --> 00:10:06,320
strike brain prompts because sometimes you might have heard actually like people say, oh I gave this

119
00:10:06,320 --> 00:10:12,000
prompt, this is not the outcome I expected from it. And couple of sessions I think I've been on,

120
00:10:12,000 --> 00:10:16,320
like I asked a lot of people, you know, like out of curiosity, you know, how many people, how many of

121
00:10:16,320 --> 00:10:21,360
them can say they write brain prompts? And it's quite hilarious because every time I see one or two people

122
00:10:21,360 --> 00:10:26,640
just raising hands. So like even today, you know, because it's it's like three, four years now, like you

123
00:10:26,640 --> 00:10:32,240
would expect like people know this. But now what happens is like a lot of people when they talk about

124
00:10:32,240 --> 00:10:37,040
AI readiness, they don't talk or just building, they also talk about how the company as a whole is moving.

125
00:10:37,040 --> 00:10:43,040
Is everyone having access to the latest tools? How those tools are being used? How can you like, you know,

126
00:10:43,040 --> 00:10:48,560
make the productivity sort of better? So you might have heard about token maxing is like one of the

127
00:10:48,560 --> 00:10:54,000
things which have been happening, I think I did some articles around it like that people are trying to

128
00:10:54,000 --> 00:11:00,080
use that and make sure they're using maximum of tokens to show productivity. So there are a lot of

129
00:11:00,080 --> 00:11:06,320
a lot of things happening. And so when you talk about AI readiness, like firstly, it's if you're

130
00:11:06,320 --> 00:11:10,960
thinking if I'm builder's lens, you should be able to ensure you use the right language model,

131
00:11:10,960 --> 00:11:16,160
you have like a right use case for it. What is the goal you are trying to achieve with that use case?

132
00:11:16,160 --> 00:11:23,120
The right tool that fits into your stack and then obviously like good quality data that you can

133
00:11:23,120 --> 00:11:27,920
feedback into it and it doesn't have to be a knowledge base. It could be historical, some

134
00:11:27,920 --> 00:11:32,640
historical data, it could be some instructions or testing data, like whatever you can leverage

135
00:11:32,640 --> 00:11:37,760
essentially. And then second part of readiness is like making sure everybody in organization has

136
00:11:37,760 --> 00:11:42,480
access to that tools and no matter where they are in the career journey, they're able to use

137
00:11:42,480 --> 00:11:48,240
that effectively for their jobs. And you know because anything which is repeated can be automated.

138
00:11:48,240 --> 00:11:54,160
So if that thought process is there, you can save more time to do those tasks which require

139
00:11:54,160 --> 00:11:58,480
like a lot more effort because I know a lot of things I've been part of there's always some technical

140
00:11:58,480 --> 00:12:03,280
depth. And you know like there's always a thing you know, oh we have to do these things but we can't

141
00:12:03,280 --> 00:12:09,440
do this because we have to prioritize these X items. So maybe like that's one arena like people can

142
00:12:09,440 --> 00:12:13,920
think about when they think about air readiness. Okay this work is something we can probably automate

143
00:12:13,920 --> 00:12:22,320
with AI and let's see if it can do the job for you. Yeah. Would you say AI can compensate poor data?

144
00:12:22,320 --> 00:12:30,800
When you say composite, what did I do mean like like if we use AI for like

145
00:12:30,800 --> 00:12:38,960
like analyzing or checking anything that it's okay? Yeah or make the data better. I think a little

146
00:12:38,960 --> 00:12:48,720
bit about the co-pilot in fabrics so it can bring the data formats in the I don't know that's

147
00:12:48,720 --> 00:12:57,520
make sense. That's something like this. But did you see AI also can help here or is it really

148
00:12:57,520 --> 00:13:07,040
human manual work to get good data? So I think with data quality, I think like I would say like

149
00:13:07,040 --> 00:13:13,360
it can help you explore the avenues. But I still I feel like I know a lot of platforms offer some

150
00:13:13,360 --> 00:13:19,200
like facilities now in which you can assist you but I think you will still have to do some

151
00:13:19,200 --> 00:13:23,520
I would say at least a part of the work to make sure that the data quality is up to the level you

152
00:13:23,520 --> 00:13:28,400
would need it to be. So I think it's gonna be more like collaboration. It wouldn't be just one

153
00:13:28,400 --> 00:13:33,040
thing or the other but if you use it in the right way it can assist you but I think if I would see

154
00:13:33,040 --> 00:13:38,160
does like a human AI collaboration other than hey can AI do it, can human do it sort of and if they

155
00:13:38,160 --> 00:13:43,520
go hand in hand together they can be as a higher chance of making things useful especially in the

156
00:13:43,520 --> 00:13:47,920
space of data quality I would say. So I don't think it's just gonna be one or the other. I think it can

157
00:13:47,920 --> 00:13:51,200
help you identify a lot of avenues. It's more like a collaboration between the two things.

158
00:13:51,200 --> 00:14:02,720
I think I don't know before I data analytics was also a big big thing. How did you

159
00:14:02,720 --> 00:14:16,000
think the analytics part has changed from these times before before pre-area and as I first

160
00:14:16,000 --> 00:14:23,760
will now? So I think it's a very good question because I data analytics is obviously like it's

161
00:14:23,760 --> 00:14:28,240
data teams especially analytics is sit with the business and their job is to support the

162
00:14:28,240 --> 00:14:34,000
business and a lot of job actually involves building reports and analyzing large data sets maybe

163
00:14:34,000 --> 00:14:40,000
writing SQL queries to fetch data from whichever background you're using. Now when we talk about

164
00:14:40,000 --> 00:14:45,440
implementing AI it's like people are moving towards conversational AI where they want AI

165
00:14:45,440 --> 00:14:50,960
assisted analytics. So essentially what I've seen is like you still use you will still use the

166
00:14:50,960 --> 00:14:57,200
same data set but then like AI would be sort of like your assistant and that would facilitate the

167
00:14:57,200 --> 00:15:01,680
outcome for you. Now when we talk about it's a building dashboard right it can assist you for example

168
00:15:01,680 --> 00:15:07,200
hey I want to write that expression and maybe like somebody is not super pro in this or something so

169
00:15:07,200 --> 00:15:13,760
they can use AI to get some support. For example like you know they can use some ideas for designing

170
00:15:13,760 --> 00:15:18,880
visualization or like so it's more or more of like you know you can build it but it will assist you

171
00:15:18,880 --> 00:15:23,920
in a larger scheme of things but then like there's another side of it like for example you want to

172
00:15:23,920 --> 00:15:28,320
analyze data and you can build an AI agent but then for that you will need the data that you will

173
00:15:28,320 --> 00:15:32,960
still use for analysis and then you will have to verify that as well unless you have automated checks

174
00:15:32,960 --> 00:15:37,680
that can do that for you. So today let's see you're on an A/B testing or a campaign data

175
00:15:37,680 --> 00:15:43,760
and you've built an AI agent for like giving sort of recommendations or I would say build a

176
00:15:43,760 --> 00:15:48,240
report for yourself which you have to present somewhere. Now okay Dran it did the entire thing

177
00:15:48,240 --> 00:15:52,320
but then still you will have to check right like what's happening unless you have built some

178
00:15:52,320 --> 00:15:58,080
automated tests because I think what the larger part of whole process is like you cannot be sure

179
00:15:58,080 --> 00:16:02,880
unless there's some sort of validation checkpoint available that the outcome is right. So

180
00:16:02,880 --> 00:16:08,720
I see conversation analytics is one area which is definitely a lot in discussion and some people

181
00:16:08,720 --> 00:16:13,680
actively working on it and then obviously like dashboard building has become like you can see

182
00:16:13,680 --> 00:16:19,040
a lot of people using AI in conjunction with dashboards to like you know build them faster and I think

183
00:16:19,040 --> 00:16:24,320
like there are other arenas as well which I know that I know and I have been part of some where like

184
00:16:24,320 --> 00:16:31,040
people are actually thinking that maybe AI can do a lot more panellitic work because it can.

185
00:16:31,040 --> 00:16:36,160
But I think like for example designing by framing. So I know a lot of people use now

186
00:16:36,160 --> 00:16:41,200
cloud design for example because it can do that for you and previously we were using

187
00:16:41,200 --> 00:16:46,000
different other tools to build by our frame so it recreates a lot of things quite quickly

188
00:16:47,360 --> 00:16:52,720
and you can for example see them in a matter of five or two seconds or maybe less than a minute or

189
00:16:52,720 --> 00:16:57,040
so yeah like there are arenas you have to identify which arena you want to use it for

190
00:16:57,040 --> 00:17:01,920
and like you know maybe brainstorming is funny arena like when you are designing dashboards.

191
00:17:01,920 --> 00:17:05,840
So conversation analytics is definitely one thing which I think is a lot of people are actively

192
00:17:05,840 --> 00:17:13,200
working on to ensure that AI can support them and you know speed up the process of you know analyzing data.

193
00:17:14,000 --> 00:17:23,360
Yeah I think since I don't know 20 years a lot of people say excel is that it looks like it's still

194
00:17:23,360 --> 00:17:31,120
existing in a lot of companies but whereas really getting still I found it's what's happened with

195
00:17:31,120 --> 00:17:36,800
Power BI. It's so hyped and now nearly no one writes about Power BI.

196
00:17:38,800 --> 00:17:48,000
I mean I think I still I mean I'm using excel so in part so I wouldn't say it's gone so I think like

197
00:17:48,000 --> 00:17:54,560
it's different perspectives I would say like I mean like Power BI is a great tool and I've used

198
00:17:54,560 --> 00:18:02,560
myself like a lot of times in the past and like I don't know like what other people's opinion are

199
00:18:02,560 --> 00:18:07,200
but I think like with the capability that the co-pilot provides you can create like great

200
00:18:08,480 --> 00:18:13,120
things on Power BI like I've seen so many examples on you know socials and LinkedIn and so on people

201
00:18:13,120 --> 00:18:19,760
publish the reports and I think like with obviously with co-pilot becoming more you know more

202
00:18:19,760 --> 00:18:24,160
a little more like keep on developing and keep on becoming a bigger product I think it will

203
00:18:24,160 --> 00:18:28,320
probably go very well along with like how power way it will become like a bigger product probably

204
00:18:28,320 --> 00:18:33,680
from where it is now so I think it's like one of those arenas I think people have different opinions

205
00:18:33,680 --> 00:18:38,720
on this so I think like I personally feel like both of those products are like I've used them quite a

206
00:18:38,720 --> 00:18:46,640
lot and I still use both of them so yeah yeah you're also an expert in the AI Microsoft AI Foundry

207
00:18:46,640 --> 00:18:54,640
as a I foundry I foundry I don't know what what what name we choose I think everyone

208
00:18:54,640 --> 00:19:02,560
I once they use another name but what what is it exactly so Microsoft Foundry yeah it was

209
00:19:02,560 --> 00:19:08,080
before it was really I think it was basically Azure AI Foundry so Microsoft Foundry is sort of like your

210
00:19:08,080 --> 00:19:13,840
enterprise AI platform so if a developer yeah engineer or a data scientist you can use it to

211
00:19:13,840 --> 00:19:18,640
build like enterprise scale applications so it offers a lot of capabilities like for example that

212
00:19:18,640 --> 00:19:23,360
tool has a model catalog so you can see all the different models it offers and then it provides sort

213
00:19:23,360 --> 00:19:28,560
of like you know breakdown by cost throughput benchmarking and so on so you will be able to see

214
00:19:28,560 --> 00:19:33,840
like which model is performing the best or sort of like which is a top performing model then it

215
00:19:33,840 --> 00:19:40,560
you can build agents guardrails memory and you can also find junior models you will definitely need

216
00:19:40,560 --> 00:19:45,600
like to deploy one of the models before you can use it and then it also offers capabilities

217
00:19:45,600 --> 00:19:50,720
on like observability and tracing so basically is an entire sort of package I would say which has

218
00:19:50,720 --> 00:19:57,600
a lot of stuck to offer and like I would say like I've used like it a lot quite a lot and like

219
00:19:58,480 --> 00:20:04,480
it is obviously like new things have been added and like it's growing like constantly and I think

220
00:20:04,480 --> 00:20:11,120
like the product also if you use it like if you like try to build it you will be able to use it

221
00:20:11,120 --> 00:20:16,560
quite easily like it's very user friendly like in the sense like you can like easily get up to speed

222
00:20:16,560 --> 00:20:21,840
to like you know build stuff and you know test out and like new features are coming in every other day

223
00:20:21,840 --> 00:20:26,080
so basically if you are a developer or even if you are like somebody who is exploding the rena

224
00:20:26,080 --> 00:20:30,160
I would recommend you to give it a try because it actually helps you to learn a lot more about like

225
00:20:30,160 --> 00:20:34,800
you know how AI can actually be implemented and you can be your own agent you can wonder your

226
00:20:34,800 --> 00:20:40,720
own agent and you can also evaluate it so you have like this entire thing in one one place which I

227
00:20:40,720 --> 00:20:50,320
think is really cool yeah and and and how well there was Microsoft fabric fit into the story

228
00:20:50,960 --> 00:20:56,320
so like basically like when you have to create like a knowledge base which is sort of like the data

229
00:20:56,320 --> 00:21:01,200
that you provide or information that you provide so you can have fabric IQ so they have there's like

230
00:21:01,200 --> 00:21:05,840
fabric IQ there is work IQ and there is found break here so they these three are sort of like they can

231
00:21:05,840 --> 00:21:11,920
work in collaboration sort of and if you are using or building ontologies or semantic layer on fabric

232
00:21:11,920 --> 00:21:17,760
you can also use that in like conjunction with found reagents as well so basically like that's where

233
00:21:17,760 --> 00:21:22,480
like you can merge sort of those things together so they are independent they can be collaborative

234
00:21:22,480 --> 00:21:32,160
you can say that I think we have another tool co-pilot studio where did you see the the difference

235
00:21:32,160 --> 00:21:41,680
to AI foundry over co-pilot studio so like like both of those tools are great but like I think

236
00:21:41,680 --> 00:21:49,360
with foundry like I have found this like I feel like I like I use it a lot and I feel like you

237
00:21:49,360 --> 00:21:53,680
with co-pilot I would say both of them are great and both of them sort of like parallel capabilities

238
00:21:53,680 --> 00:21:59,360
but like using it for example building foundry IQ for me personally has been really

239
00:21:59,360 --> 00:22:06,720
easy and I would say really like useful and because a lot of my like building those applications

240
00:22:06,720 --> 00:22:12,080
involves like using data effectively and like in a larger sense so I think like and also understanding

241
00:22:12,080 --> 00:22:14,960
those models it gives me a lot more capabilities to use those things.

242
00:22:14,960 --> 00:22:27,120
What did you think is is AI foundry also for smaller mid-sized companies or is it an enterprise

243
00:22:27,120 --> 00:22:41,360
product? Okay one second give me one second. So I think like Microsoft boundaries obviously more

244
00:22:41,360 --> 00:22:46,240
enterprise AI and it basically helps to ensure that like it creates your gives you all the

245
00:22:46,240 --> 00:22:53,120
capabilities that you would need and like when we talk about like you know for example when you

246
00:22:53,120 --> 00:22:58,240
build the foundry IQ you can also implement Azure AI search in the backend or you can also implement

247
00:22:58,240 --> 00:23:04,480
for example things like you know work IQ and also like you can also create additional capabilities

248
00:23:04,480 --> 00:23:09,920
on that. So I think if you try building it like see there are two things is tools and then there is

249
00:23:09,920 --> 00:23:14,880
sort of like your knowledge base. So the tools section you can only upload files so it's quick

250
00:23:14,880 --> 00:23:19,440
and faster but if you're building enterprise agent then you will need sort of like more solid

251
00:23:19,440 --> 00:23:24,400
knowledge base and in those capabilities it's sort of important to ensure that like you have a

252
00:23:24,400 --> 00:23:29,440
larger set of data you can engineer the data effectively and make sure that like that data can be

253
00:23:29,440 --> 00:23:36,640
used within your application effectively. So no matter you use for example one or the other as long

254
00:23:36,640 --> 00:23:41,680
as you have like build the ontologies semantically effectively that should be able to plug in into

255
00:23:41,680 --> 00:23:46,400
the main tool so that essentially makes your job a lot more easier but yeah that is a process you

256
00:23:46,400 --> 00:23:50,320
know building that thing you will have to know what your data is and what you want to feed versus what

257
00:23:50,320 --> 00:24:00,160
you don't want to feed. And what kind of models are available and they have found? I would say there's

258
00:24:00,160 --> 00:24:06,400
a lot I think yesterday I saw you know probably over 180 or something models so it has basically

259
00:24:06,400 --> 00:24:13,680
from cloud to cloud opus, clouds on it it also has fable which was which basically came in

260
00:24:13,680 --> 00:24:21,200
co because of well all the node reasons it has grok and then like mistral and basically all the

261
00:24:21,200 --> 00:24:28,320
latest models all the GPD models it has 4.154O everything basically but I think like see the thing is

262
00:24:28,320 --> 00:24:31,840
like a lot of people think they have to use the latest model to build their applications but I think

263
00:24:31,840 --> 00:24:36,800
it's super important to understand that you don't necessarily need all the time latest model if you're

264
00:24:36,800 --> 00:24:42,160
going to find unit or if you're going to build like say a rag for it right so you have to make sure

265
00:24:42,160 --> 00:24:48,080
that what's a cost are you willing to pay for an application and decide then which one you want to

266
00:24:48,080 --> 00:24:52,640
use to build it and then the second thing you would want to also make sure is like what data

267
00:24:52,640 --> 00:24:57,200
will you use to find unit or build the rag application. I think that's important I know the

268
00:24:57,200 --> 00:25:02,480
example because I was saying rag is dead because of like larger context with was available but I think

269
00:25:02,480 --> 00:25:08,560
it's yeah I mean if unless you have a better approach then yeah you can you can debate about it.

270
00:25:09,520 --> 00:25:18,320
Yeah I think a lot of people now yeah they're using AI, using AI but we also have this

271
00:25:18,320 --> 00:25:26,240
Azure machine learning stuff and I don't know what to search makes sense especially for

272
00:25:26,240 --> 00:25:33,600
antiprizes to look where I spent I think expensive tokens and where can I use other things

273
00:25:34,240 --> 00:25:44,800
like that. Okay so what did you think how will the AI found really evolving over the next month?

274
00:25:44,800 --> 00:25:51,600
I mean there's I mean it's constantly evolving like obviously I think the new models will keep

275
00:25:51,600 --> 00:25:55,680
up from coming so that's gonna be always there for sure so I think like that's one thing which

276
00:25:55,680 --> 00:26:02,640
is going to be there then I think memory management because you know like the or agents cannot work

277
00:26:02,640 --> 00:26:08,720
effectively so I like that's something which probably will come along the way because like that

278
00:26:08,720 --> 00:26:13,280
will really I would say that's really the game changer as well in a way because if your AI does not

279
00:26:13,280 --> 00:26:17,920
remember anything about a previous conversation then I think it would probably be one of those things

280
00:26:17,920 --> 00:26:24,160
like it resets every single time so it does not adapt episode of so it requires to remember all those

281
00:26:24,160 --> 00:26:29,040
things that you're basically asking it previously so let's say today you are trying you're working in

282
00:26:29,040 --> 00:26:35,360
public sector you want to pay your council tax and maybe like you are a person and if your details

283
00:26:35,360 --> 00:26:40,720
are not provided it cannot give you the like account details or how to pay for example your tax

284
00:26:40,720 --> 00:26:45,920
or do you want to refund for that tax with super important to understand that that would be something

285
00:26:45,920 --> 00:26:53,920
which can actually very change it. I think it's a big part of also context engineering which was very

286
00:26:53,920 --> 00:27:00,800
at the end the term came way back in 2023-24ish start I would say but now it has just grown a lot and

287
00:27:00,800 --> 00:27:06,320
like it has become a super important part and I've also heard a lot about like harness engineering as

288
00:27:06,320 --> 00:27:11,680
well which is like something which people are discussing right now and I think like this this is

289
00:27:11,680 --> 00:27:19,360
one of the arena which I think definitely is going to grow as time goes by. And what role

290
00:27:20,160 --> 00:27:27,600
plays security and compliance inside the I foundry or is it inside the F I foundry?

291
00:27:27,600 --> 00:27:33,440
So obviously you can build guardrails so like I've seen people talking a lot about guardrails but

292
00:27:33,440 --> 00:27:38,480
like a lot of people they don't necessarily understand or know about what guardrails really mean so

293
00:27:38,480 --> 00:27:44,480
guardrails basically protect and make sure that like your agent does not like you know it doesn't

294
00:27:44,480 --> 00:27:51,840
have any unintended behavior right that's really the purpose of guardrails and so by default it

295
00:27:51,840 --> 00:27:56,560
foundry there are some guardrails in build but you can always equate any one so for example

296
00:27:56,560 --> 00:28:03,760
like if your application does not require any specific kind of content or protected material or

297
00:28:03,760 --> 00:28:10,160
something like sensitive data sort of so it helps to ensure that like okay this data this particular

298
00:28:10,160 --> 00:28:15,040
application should not bypass these things it should not we find a way to like tackle those things

299
00:28:15,040 --> 00:28:21,440
or for example you have fed some like the IID data which I mean most of us will not but like say

300
00:28:21,440 --> 00:28:26,720
hypothetically you did then how will like manage it so I think guardrails ensure that there is like

301
00:28:26,720 --> 00:28:32,000
you know those lanes that should be controlled and that your application does not perform any

302
00:28:32,000 --> 00:28:36,880
unintended behavior it is an option to also create redeeming so you can like create a redeem and

303
00:28:36,880 --> 00:28:42,640
sort of your test stress test your agent in a way that you can create scenarios and test it when

304
00:28:42,640 --> 00:28:46,880
it's gonna break one thing I always say is like you know we should perform like people should

305
00:28:46,880 --> 00:28:51,280
perform as much as negative testing when they build these things simply because what it does for

306
00:28:51,280 --> 00:28:58,400
you is like any situation that you might think can happen it should exactly behave in a way it should

307
00:28:58,400 --> 00:29:04,240
like it should not be saying oh this is yes but actually the answer is no for it so like it makes

308
00:29:04,240 --> 00:29:08,560
sure that you are actually active you're trying to test it without waiting for somebody else to

309
00:29:08,560 --> 00:29:14,080
tell you so imagine you build a product right you are a company and you go to a customer and they

310
00:29:14,080 --> 00:29:19,840
test it and you find those flaws for you then you they might not be convinced even if you fix those

311
00:29:19,840 --> 00:29:24,800
flaws effectively because now in the mind you're like oh I can't I don't know if this is gonna be like

312
00:29:24,800 --> 00:29:30,080
something I want to invest in so it's a proactive attitude also like check those things and

313
00:29:30,080 --> 00:29:35,760
just as many scenarios to make sure that okay I've done this my new diligence on this thing I know

314
00:29:35,760 --> 00:29:40,880
that these scenarios is gonna fail the scenarios gonna work so you already know all the pros and cons

315
00:29:40,880 --> 00:29:45,360
of your product and then let's say you know how to rectify them eventually because otherwise you

316
00:29:45,360 --> 00:29:50,480
will not identify them so yeah I think like this is one of the other interesting area where you can

317
00:29:50,480 --> 00:29:57,120
you have to practice like as it is possible AI where like you know just building something but

318
00:29:57,120 --> 00:30:07,600
also building something which is effective useful I think we spend a lot of I don't know money in

319
00:30:07,600 --> 00:30:16,480
AI and in tokens and in building AI also security and all all the other costs that are often

320
00:30:17,360 --> 00:30:27,360
done can is the CFO and ask yeah what is the the return of invest if they are or have you any tips

321
00:30:27,360 --> 00:30:35,040
how company can yeah elevate the every from AI investments that's a good question I think this is

322
00:30:35,040 --> 00:30:39,280
like a lot of people are discussing about just now as we I'm sure like somebody would have already

323
00:30:39,280 --> 00:30:46,000
little bit of like they imposed on the service fee I think yeah see air investment everybody wants

324
00:30:46,000 --> 00:30:52,160
use AI because they know that this is something which is gonna make like a jobs maybe like easier or

325
00:30:52,160 --> 00:30:59,040
like make us more productive and so on when we talk about using AI a lot of times people don't know

326
00:30:59,040 --> 00:31:05,440
where they want to use AI like which which this problem am I trying to solve with is AI as a tool

327
00:31:05,440 --> 00:31:12,880
available to me like this last language model because like you are a company which has a big customer

328
00:31:13,520 --> 00:31:17,840
interaction agent centers something like that and like probably your job is to make sure that

329
00:31:17,840 --> 00:31:24,160
the basic queries is supported by AI and then like maybe the very very complicated and once

330
00:31:24,160 --> 00:31:30,560
where is it users are asking then maybe I will do my agent it let's that's your case now you are

331
00:31:30,560 --> 00:31:36,880
trying to save the time but call me like how do you how that's basically where you have to see

332
00:31:36,880 --> 00:31:42,240
where you can solve the problem and how do you make sure that like AI is effectively useful because

333
00:31:42,240 --> 00:31:47,200
it could be possible that one of those cases is not the right use case for very one

334
00:31:47,200 --> 00:31:53,040
implement AI so I think it started identifying the right use case first thing and then understanding

335
00:31:53,040 --> 00:31:59,040
like what's my current cost that I'm paying for that same workflow or process versus what I think AI

336
00:31:59,040 --> 00:32:05,520
can do for me and then start being a little bit real about like comparison that okay for example

337
00:32:05,520 --> 00:32:10,880
I am paying X amount of pound million pounds here if I don't pay this I move this investment here

338
00:32:10,880 --> 00:32:16,560
will that give me RWA in the long term and it should also be thought of like you know not just like

339
00:32:16,560 --> 00:32:21,920
oh RWA is gonna come in like one month or two months maybe like have a plan of like in how much time

340
00:32:21,920 --> 00:32:25,680
do you think you can recover that money for example it's not like you know in one month you're gonna

341
00:32:25,680 --> 00:32:31,040
get the entire RWA right like it depends on like how big or small investment is in that

342
00:32:31,040 --> 00:32:36,000
sure and I think like then obviously like that from that like then starts like your

343
00:32:36,000 --> 00:32:41,520
for a building phase so in that start building phase you have to make as I said like a robust design

344
00:32:41,520 --> 00:32:47,760
which will cater to that investment that you are building and from there you have to start like

345
00:32:47,760 --> 00:32:53,120
being like okay now I can see that there is a possibility that this AI will let's say we 30%

346
00:32:53,120 --> 00:32:59,120
of my investment that I was doing in the process A versus process B so I think like there is like

347
00:32:59,120 --> 00:33:05,840
the thing is like RWA will come in at a point in time so I think in my understanding whether

348
00:33:05,840 --> 00:33:11,200
RWA will come is also the key it's not gonna be like in one week I'll get all the you know

349
00:33:11,200 --> 00:33:15,920
this value back or maybe in one month so that's thing and then identifying the right case I think

350
00:33:15,920 --> 00:33:21,120
these are the four or five things are super important when you are in organization trying to implement AI

351
00:33:21,120 --> 00:33:26,240
and then yeah also like also having a great like the right team because having the right people

352
00:33:26,240 --> 00:33:30,320
who are building the right architecture is important because you would know right like for example

353
00:33:30,320 --> 00:33:34,560
if you're a data platform architect right if you design the right architecture for example you

354
00:33:34,560 --> 00:33:39,840
are catering to five different products right in some like some areas the latency or the cost is

355
00:33:39,840 --> 00:33:45,520
more important in some cases real-time streaming data is important and you cannot have same thing

356
00:33:45,520 --> 00:33:49,280
or the other you have to think about each in every section so this is the same thing when we talk

357
00:33:49,280 --> 00:33:55,440
about AI like your architecture the design and then like whatever investment you're gonna make in

358
00:33:55,440 --> 00:34:02,400
that entire product or an application you are building is really the key to make sure what

359
00:34:02,400 --> 00:34:07,920
problem it will solve today you decide it's gonna be for sort of like encoding assistant for you

360
00:34:07,920 --> 00:34:13,520
or some sort of like I don't know like maybe assistant that will help you make

361
00:34:13,520 --> 00:34:20,880
DevOps tickets or like for example G-RT tickets or something then watch my current effort versus

362
00:34:20,880 --> 00:34:27,200
what's the effort now then what's the time I'm investing now what's the time after implementing AI

363
00:34:28,240 --> 00:34:33,120
how much money was going into this product before versus how much money do I plan to recover in

364
00:34:33,120 --> 00:34:39,600
one year three years time and this is how they should go about it rather than like expecting that

365
00:34:39,600 --> 00:34:43,760
this will happen just in one go and also it's an improvement process like you have to keep

366
00:34:43,760 --> 00:34:48,960
on improving the application a lot of times it's just sitting it's built in next year but like you

367
00:34:48,960 --> 00:34:54,640
know live language models they have to be monitored because sometimes they can change they can

368
00:34:54,640 --> 00:34:58,800
hallucinate even after you've done everything well so you have to keep them monitoring and tracing

369
00:34:58,800 --> 00:35:02,720
how they're performing and then if you think that at some point there is something else you can do

370
00:35:02,720 --> 00:35:07,120
and make it better because you know with context that you provide to the models right like you have

371
00:35:07,120 --> 00:35:12,320
to expand the old data out if you don't use it otherwise it's gonna be challenging for somebody to

372
00:35:12,320 --> 00:35:17,520
like check every single aspect of that application so I think there's so many parts of also building

373
00:35:17,520 --> 00:35:22,080
that product once you are entering that phase that the product team has to ensure

374
00:35:24,240 --> 00:35:29,120
to make sure that it performs in the way it is intended to perform and that's basically where

375
00:35:29,120 --> 00:35:33,600
it starts generating R-A because it's a whole cycle I would say from start to end.

376
00:35:33,600 --> 00:35:40,240
Yeah you say hallucination I think for me I think the

377
00:35:40,240 --> 00:35:46,720
I see the model is getting better better but now we are also in the world of AI agents where

378
00:35:46,720 --> 00:35:55,680
one agent said the other agent did this and so on so a lot of I'd say specialists I think that's

379
00:35:55,680 --> 00:36:05,200
experts I think it's come from mistrial the idea to build different experts but how how

380
00:36:05,200 --> 00:36:11,200
many chore can be monitored the outcomes that we have no bias or hallucinations in the

381
00:36:11,200 --> 00:36:19,200
work-restance. So like when you talk about hallucinations right like I know the latest model that we

382
00:36:19,200 --> 00:36:24,960
have gotten from like different all the different AI companies I would like I would say they are

383
00:36:24,960 --> 00:36:29,280
slightly better but obviously it's still there the hallucination is still there now people use

384
00:36:29,280 --> 00:36:33,760
temperature settings they have like different settings available that it can use to control the

385
00:36:33,760 --> 00:36:39,440
hallucinations but I think I would say like when you test your application there should be a

386
00:36:39,440 --> 00:36:45,760
baseline what's the acceptable behavior for my application and like what's the intended behavior

387
00:36:45,760 --> 00:36:50,880
so if my agent is going to hallucinate and give me deviation like for example with 5-10%

388
00:36:50,880 --> 00:36:55,920
then I would probably say it's not ready for production because I would not be able to accept

389
00:36:55,920 --> 00:37:01,920
that simply so I think like when you build applications especially with hallucinations it should

390
00:37:01,920 --> 00:37:06,720
be like different cases where you would have tested those hallucinations how much you are getting

391
00:37:06,720 --> 00:37:12,400
those hallucinations for then what's the baseline if it's like for example you know confidence

392
00:37:12,400 --> 00:37:16,960
code because after confidence code of your hallucination is like slight deviation which is probably

393
00:37:16,960 --> 00:37:24,560
would say like maybe like 0.0001 sort of or like very small which is not so much and also the answer

394
00:37:24,560 --> 00:37:28,640
is correct I think that's like one thing which you have to make sure you are checking every single

395
00:37:28,640 --> 00:37:33,040
time and then like maybe test a lot of times to see like what's the baseline you're going to get

396
00:37:33,040 --> 00:37:37,760
what's acceptable percentage you can also create like confidence scores for example if my

397
00:37:37,760 --> 00:37:44,560
application every time gives me 90% of confidence which I know is going to be right and then also automate

398
00:37:44,560 --> 00:37:51,360
some like checks in every step like not like in that conversation is in every turn like every time

399
00:37:51,360 --> 00:37:58,240
it interacts that's also I think very important and then like if it falls below that baseline then okay

400
00:37:58,240 --> 00:38:02,560
now we need to see so they should be like process for that so I think in that process you can make

401
00:38:02,560 --> 00:38:06,720
it semi-automate but then somebody will have to still check there should be some human in the loop

402
00:38:06,720 --> 00:38:11,600
especially if you're working for like industries which are very very you know sensitive like for

403
00:38:11,600 --> 00:38:15,440
example health care so you have to make sure somebody is checking those things constantly because

404
00:38:15,440 --> 00:38:23,040
yeah like you're I would say like like you cannot go wrong with like your outcomes I mean I'm like

405
00:38:23,040 --> 00:38:27,040
always like advocating for this like you cannot go wrong with what it creates we have to make sure

406
00:38:27,040 --> 00:38:32,160
that we're always checking those things step by step in every process so build evaluations check

407
00:38:32,160 --> 00:38:38,000
maybe like relevance accuracy how core and the result is is it having task adherence which means

408
00:38:38,000 --> 00:38:43,440
like it does that task it's a boost you do and then maybe give also some sort of feedback like sort

409
00:38:43,440 --> 00:38:48,320
of a feedback loop so that the limb tries to improve itself and like you know like also tries to

410
00:38:48,320 --> 00:38:52,800
find better ways to solve the same problem maybe look if you have given a better prompt it's structured

411
00:38:52,800 --> 00:38:57,840
properly I think these things sort of like help to understand where the hallucination is happening

412
00:38:57,840 --> 00:39:02,000
from and then how can we rectify it it's not just about the settings it's also about how you

413
00:39:02,000 --> 00:39:13,360
tackled outcomes once you get it. I think now we see I think every product now includes AI

414
00:39:13,360 --> 00:39:25,200
at home my my vacuum my my my coffee machine all has now AI in it and last of the in-depth that's

415
00:39:25,200 --> 00:39:31,680
a new new buzzword I see monitor this it's was responsible AI I don't know what it do but

416
00:39:31,680 --> 00:39:41,280
responsible I it's yeah it's as a new LinkedIn it's everywhere what do the actual mean in practice.

417
00:39:41,280 --> 00:39:50,080
Responsibility I responsibility I basically like you are using like the AI applications in a way

418
00:39:50,080 --> 00:39:56,640
that like you know way that it's actually fulfilling the task that you have assigned it to do and

419
00:39:56,640 --> 00:40:02,400
if it at some point like for example initially when chat you be decay I know a lot of people just

420
00:40:02,400 --> 00:40:08,560
fed in data or gave it information because they thought like you know just take all the information

421
00:40:08,560 --> 00:40:13,520
and that's it but you cannot put like any other data random data into you know applications because

422
00:40:13,520 --> 00:40:19,440
you know there's a large language models so using AI effectively is always important to ensure

423
00:40:19,440 --> 00:40:27,840
that like when you're getting a response you so if you are a user like for example if I'm the user

424
00:40:27,840 --> 00:40:32,480
right let's say I'm expecting a response in some capacity I would know that is this a response

425
00:40:32,480 --> 00:40:36,720
right or wrong I mean how are people how do people know that AI is getting wrong response if they're

426
00:40:36,720 --> 00:40:41,840
not the subject matter experts of their particular area right so understanding if I can rely on

427
00:40:41,840 --> 00:40:47,520
trust on this and on this outcome so like if you are generating some content for example

428
00:40:48,640 --> 00:40:53,200
you're using it from brainstorming do you think the large idea is going to be practical and like

429
00:40:53,200 --> 00:40:59,680
for example if you're asking on seeking information are those links reliable are the right links or

430
00:40:59,680 --> 00:41:05,680
that information is right so I think I recently tested one of the example like I was trying to build

431
00:41:05,680 --> 00:41:11,040
one agent for like you know just being my concert tax right and I was trying to see like how will it

432
00:41:11,040 --> 00:41:15,440
like for example take the data so it gave me some made a website you know in the start and I was like

433
00:41:15,440 --> 00:41:19,680
oh this link doesn't meet me anywhere because that's not the right website even and the information

434
00:41:19,680 --> 00:41:25,440
was incorrect so measuring and comparing that this is the outcome I'm getting this is the realistic

435
00:41:25,440 --> 00:41:29,920
outsmash should have been getting is making sure that you're responsibly using it especially

436
00:41:29,920 --> 00:41:35,440
like not just trusting whatever it's been given because even on both start ones is written that it

437
00:41:35,440 --> 00:41:41,040
makes an incorrect responses on the bottom of that the bar where you give the prompt so I think

438
00:41:41,040 --> 00:41:47,200
it starts with understanding that you have to like make sure you are checking everything before using it

439
00:41:47,200 --> 00:41:53,360
and like if it's not like sort of like an organizational application if you're especially

440
00:41:53,360 --> 00:41:57,920
using for example chatchipity or any other yeah I tool available in the market which this

441
00:41:57,920 --> 00:42:04,720
is a lot of them now and I think like with using a to for example you phone on a daily life first

442
00:42:04,720 --> 00:42:11,600
stuff so if you're building let's say you're planning a travel item right that's the one the muskony

443
00:42:11,600 --> 00:42:19,440
skills I've heard people using it for like making sure that like like it gives you the right option

444
00:42:19,440 --> 00:42:23,840
of the flight links and then checking those flight links actually available it basically starts

445
00:42:23,840 --> 00:42:28,640
from all those small steps and then obviously like from enterprise level you all know that like we

446
00:42:28,640 --> 00:42:34,560
need more testing and more like evaluation evaluations and LLM as a judge and you know making sure

447
00:42:34,560 --> 00:42:39,200
that all those things work hand in hand to make sure it's more responsible and then like for example

448
00:42:39,200 --> 00:42:43,360
if you're even if you're working in an organization if you find that some points giving wrong answers

449
00:42:43,360 --> 00:42:47,600
please share with other people as well so that they are more aware as well because sometimes maybe

450
00:42:47,600 --> 00:42:52,000
you identify a use case and other people haven't come across that so they will also be aware that

451
00:42:52,000 --> 00:42:56,240
oh like this gives what potentially happened like I couldn't get a wrong outcome for this thing

452
00:42:56,240 --> 00:42:59,760
so I think that starts with you know sharing more knowledge and then obviously like you know

453
00:42:59,760 --> 00:43:03,840
these like we are having for example community meetups as well so a lot of people come and share

454
00:43:03,840 --> 00:43:08,240
their projects and they share like things they have found so I think also listening to other people

455
00:43:08,240 --> 00:43:12,720
and like you know learning from them along the way also helps because for example they might have

456
00:43:12,720 --> 00:43:17,520
come across use case which you haven't so you were also know now okay maybe I will think about

457
00:43:17,520 --> 00:43:22,400
this next time and I'm building something new for me that I should be I might come across the same

458
00:43:22,400 --> 00:43:27,520
scenario and I will try to sort of not fall in that space where I have to do the like work to like

459
00:43:27,520 --> 00:43:31,600
you know reinterpret it so I think it starts with multiple practices to make sure that you are

460
00:43:31,600 --> 00:43:42,000
using AI responsibly yeah actually a lot of companies or selling AI governance frameworks there

461
00:43:42,000 --> 00:43:48,240
there are a lot of it did company really need an AI governance framework or it's okay to have

462
00:43:48,240 --> 00:43:55,120
IT in the data governance framework what's your opinion here I think oh AI governance should in like

463
00:43:55,120 --> 00:43:59,680
in the granting of things AI governance should be like big things since the beginning of the process

464
00:44:00,400 --> 00:44:05,200
because I think it's sometimes like we build the whole thing and then we think about governance

465
00:44:05,200 --> 00:44:10,720
I would say I think it should be in the start and it's actually one of the reasons for your AI

466
00:44:10,720 --> 00:44:17,920
being truly ready because like if you're like all the different processes that are important for

467
00:44:17,920 --> 00:44:24,560
example you work for a company which is super like sensitive, user sensitive data or legal or legal

468
00:44:24,560 --> 00:44:30,240
data for example NDs and so on so like it's super important that you have taken care of the governance

469
00:44:30,240 --> 00:44:35,840
because like it would be quite surprising if somebody is going to go and use it you know because like

470
00:44:35,840 --> 00:44:40,240
those kind of information and like you know you haven't taken care of the ethics and like you know

471
00:44:40,240 --> 00:44:46,160
make sure that all the sensitive data is tackled properly I think it's like a lot of importance

472
00:44:46,160 --> 00:44:54,160
there is on that so I think you you can use like you should make it a part of the architecture itself

473
00:44:54,160 --> 00:44:58,800
in the beginning of the process it should not be an afterthought that we do it once we have built a

474
00:44:58,800 --> 00:45:04,880
product it should be big in the beginning of the process and like I think like if you have a B2B or

475
00:45:04,880 --> 00:45:14,000
B2C company like like without governance I I'm not sure if like people will be like convinced you

476
00:45:14,000 --> 00:45:20,160
buy something because they have to pass the governance checks like ensure that the product is truly

477
00:45:20,160 --> 00:45:26,160
ready for like you know being used by the people so yeah I think like it's super important I mean

478
00:45:26,160 --> 00:45:31,520
framework wise like yeah I mean a lot of companies do sell a governance frameworks it depends on

479
00:45:31,520 --> 00:45:36,240
each organization if they actually use it I mean I would say like it's best to have it built in

480
00:45:36,240 --> 00:45:40,640
the start and to see like if actually that framework is going to be what file or like use before your

481
00:45:40,640 --> 00:45:45,440
product or the company that it works for and like if you have for example your own internal

482
00:45:45,440 --> 00:45:50,480
governance teams AI governance teams for example then they can support you in that process so it depends

483
00:45:50,480 --> 00:45:54,400
if you don't have internal governance teams you can support you then you might see external health

484
00:45:54,400 --> 00:45:58,640
but it totally depends on one company to another so I don't think one size would for all I think

485
00:45:58,640 --> 00:46:03,360
it depends from case to case cases those like smaller companies might I mean might not have a big

486
00:46:03,360 --> 00:46:08,000
governance team so they might need external support to do that but some companies do have a big

487
00:46:08,000 --> 00:46:13,680
governance team so they can support them in those processes yeah um what did you think about

488
00:46:13,680 --> 00:46:18,880
regulations I think we have here also two teams we have the the people they say it's allowing

489
00:46:18,880 --> 00:46:26,720
the innovation and the other saying it's an A-blig trust what what did you think uh I mean

490
00:46:26,720 --> 00:46:34,880
yeah I mean there's some industries where like uh it's important I mean it's sort of in a way

491
00:46:34,880 --> 00:46:40,240
also trying to ensure that whatever you are creating is the level as I said like entrustworthy

492
00:46:40,240 --> 00:46:47,520
so it's two sides of the same coin so it's all but I think like if your industry has this thing

493
00:46:47,520 --> 00:46:52,000
you know where they have to pass all these criteria then I think like you have to go through it

494
00:46:52,000 --> 00:46:57,280
and it's better to be like you know check all those things beforehand like and you know rather than

495
00:46:57,280 --> 00:47:04,080
like just putting something out like innovation is also go leave and like you know it's

496
00:47:04,720 --> 00:47:09,600
it's responsibly used and you know save to use also so I think yeah there's two sides of the same

497
00:47:09,600 --> 00:47:14,000
coin I would say so yeah I think some industries will definitely need to ensure that they pass all

498
00:47:14,000 --> 00:47:19,040
those barriers before they actually sort of reach a point where they are directly ready to be

499
00:47:19,040 --> 00:47:28,160
used by consumers and yeah this is awesome and how important that you think is it's actually

500
00:47:28,960 --> 00:47:36,080
yeah this is community in these AI area oh communities I mean there are lots

501
00:47:36,080 --> 00:47:42,800
now uh I mean communities are important and I think uh it's important to hear also like what other

502
00:47:42,800 --> 00:47:48,320
people are bringing like for example building because I think like collaboration and all the

503
00:47:48,320 --> 00:47:53,200
shittings are the best way to learn the new tech or any tech essentially and I think like a lot of

504
00:47:53,200 --> 00:47:58,160
places I have attended events or like I've learned a lot of things and the people I've created

505
00:47:59,040 --> 00:48:04,160
you know also like it also sparked my interest in the sense like you can discuss with those people

506
00:48:04,160 --> 00:48:09,040
what approaches they are followed so it does create a platform for people to upscale so for example

507
00:48:09,040 --> 00:48:12,160
somebody comes from core engineering like you know software engineering and then

508
00:48:12,160 --> 00:48:17,680
now they know about coding they know how to build software architecture as you see everything

509
00:48:17,680 --> 00:48:23,040
I listen so on but maybe like I want to understand more about AI which is more about like you know

510
00:48:23,040 --> 00:48:28,400
generative AI like large language models how they are built how the how they process information so

511
00:48:28,400 --> 00:48:32,640
now like obviously there are all of courses online but then like community meetups provide that

512
00:48:32,640 --> 00:48:38,160
sort of like a center spot where you can go and I generally say to people like you don't have to go

513
00:48:38,160 --> 00:48:43,200
to like you have to pick those events where you can learn the most because every event has

514
00:48:43,200 --> 00:48:47,680
their different things like some events like there's advanced AI some places they are more like

515
00:48:47,680 --> 00:48:53,600
topics which are like sort of like in middle like and some barely early start you know they have

516
00:48:53,600 --> 00:48:57,840
like topics which are very so you pick which one you are really trying to learn or maybe your job

517
00:48:57,840 --> 00:49:03,200
whichever you are in where it is more relevant and then you can go and attend those events and learn

518
00:49:03,200 --> 00:49:08,320
from them and then ask questions because curiosity is something which is super important especially

519
00:49:08,320 --> 00:49:12,240
in today's world to learn and understand more you should ask the questions if you have any

520
00:49:12,240 --> 00:49:17,360
confusion or doubt about something so that you learn along the way so I think community meetup

521
00:49:17,360 --> 00:49:22,400
are really really important in my opinion because they provide that free space to you know

522
00:49:22,400 --> 00:49:28,160
really be like learn from other people and I think either you are an attendee or you speak or like

523
00:49:28,160 --> 00:49:33,120
you're an organizer you will learn a lot because you will get different diverse perspectives

524
00:49:33,120 --> 00:49:37,680
and like if I'm a data scientist I will come from a data science lens or maybe somebody else who

525
00:49:37,680 --> 00:49:42,000
is a software engineer they will come from a software engineering lens and then somebody maybe is

526
00:49:42,000 --> 00:49:46,400
in analytics they will come from analytics lens everyone is using different tools and it's going

527
00:49:46,400 --> 00:49:51,120
to be everywhere so it's nice to know what's happening so for example in analytics we are seeing

528
00:49:51,120 --> 00:49:56,320
conversationally I being like one of the topics being discussed or in software engineering it's more

529
00:49:56,320 --> 00:50:03,760
like oh hey like now we can use AI to build code or make coding assistance sort of and if you are

530
00:50:03,760 --> 00:50:08,720
data scientist you would think for example oh like I use ML ops now maybe I'm going to use LLM ops

531
00:50:09,600 --> 00:50:14,400
or I use large language models to you know reproductive modeling how it can support there or if you're

532
00:50:14,400 --> 00:50:20,960
in finance or like marketing or sales how can I use AI for example will AI be able to help me create

533
00:50:20,960 --> 00:50:27,280
reports create campaigns analyze my campaign data so I think the different you know lenses so it's

534
00:50:27,280 --> 00:50:32,880
good to know all the lenses and learn how people are using it and if you have something to share

535
00:50:32,880 --> 00:50:35,680
you should definitely go and share that information as well.

536
00:50:35,680 --> 00:50:42,720
And Shavani you also lead a community the women in data are long then can you tell a little bit

537
00:50:42,720 --> 00:50:50,000
about this? Yeah so women in data it's a non-profit organization and it started back in 2015

538
00:50:50,000 --> 00:50:56,000
in the United States and we have this chapter which is now I think almost 900 members so being

539
00:50:56,000 --> 00:51:01,280
a part of the community for a long time and I lead the chapter I think I organized 35 or 40 events

540
00:51:01,280 --> 00:51:07,280
so far in the past few years so the community and obviously just gone by 40-50% now so we have a lot

541
00:51:07,280 --> 00:51:12,240
of different women from the years you know walks of life and the organizer banks we helped them

542
00:51:12,240 --> 00:51:17,760
and dirt like fine rolls for example and we have a team chapters in click 57 countries or something

543
00:51:17,760 --> 00:51:22,160
now and obviously this is one of the chapters biggest chapters in Europe that one we have right now

544
00:51:22,160 --> 00:51:29,040
in Gijuke yeah it's a super collaborative fun place to be like anyone can join like there's like

545
00:51:29,040 --> 00:51:34,000
anybody like irrespective of is there like been on and they can like you know be a part of the

546
00:51:34,000 --> 00:51:38,480
community they can share their sessions they can also mentor other people they are also organizing

547
00:51:38,480 --> 00:51:43,680
a data thorn in the next couple of months so if somebody wants to you know build a team and participate

548
00:51:43,680 --> 00:51:49,280
I'm sure there's gonna be some AI topic that I think the topic is what's cooking this time so it's

549
00:51:49,280 --> 00:51:52,560
gonna be on some like data will be I think on cooking but then you create your own thing

550
00:51:52,560 --> 00:51:57,280
and there's bunny projects as well so somebody wants to you know team of it someone and build a

551
00:51:57,280 --> 00:52:02,240
project there's like that option as well so there's a lot of stuff it offers and yeah it's like

552
00:52:02,240 --> 00:52:08,720
it's not profit so yeah you can just join it for free and like you have reached the the value of

553
00:52:08,720 --> 00:52:14,560
like what it offers and I think you have also another community it's the build us foundry

554
00:52:14,560 --> 00:52:23,360
yeah so yeah bill is foundry I started this year and we already have like 150 plus members I

555
00:52:23,360 --> 00:52:29,360
mean we only done I think two or three events this year because we wanted to like build high quality

556
00:52:29,360 --> 00:52:35,440
events so we do we are building ones on one and the other and one other I'm doing this with another

557
00:52:35,440 --> 00:52:41,040
MVP so we are like doing and building this community together it's mostly focus on software engineering

558
00:52:41,040 --> 00:52:47,360
architecture AI and like you know startups because you know startups there's a lot more startups in

559
00:52:47,360 --> 00:52:53,520
London out than they were I think before especially because of AI so we are like looking for people who

560
00:52:53,520 --> 00:52:58,000
are like you know who want to share their real world expertise or knowledge I mean it's okay if

561
00:52:58,000 --> 00:53:01,840
something like didn't work that was just fine we just want to know like what's working what's not

562
00:53:01,840 --> 00:53:08,640
working and it's practitioner led community so we are looking people who are more like practitioners

563
00:53:08,640 --> 00:53:15,520
in their area and so we did like I think the launch event back in April and then we did one last

564
00:53:15,520 --> 00:53:22,000
month and then we have on one coming in September okay awesome yeah let's jump in the rapid

565
00:53:22,000 --> 00:53:29,920
fire round you say well the first thing that comes in your mind so coffee tea or energy drink

566
00:53:29,920 --> 00:53:40,800
during developed agents okay the third one energy drinks okay power be our our excel

567
00:53:43,360 --> 00:53:52,960
both pie Thor as KL oh that's just I think both because I use both of them so

568
00:53:52,960 --> 00:54:04,720
probably a year or fine tuning I mean it's case based so can't pick one or the other but like if

569
00:54:04,720 --> 00:54:09,600
I want to go for one I might if it's simpler I will go for fine tuning if it's not like a simple

570
00:54:09,600 --> 00:54:15,360
use case I might go with the other one yeah your biggest productivity hack focus

571
00:54:15,360 --> 00:54:23,760
fire bridge Microsoft product favorite Microsoft product oh there are so many I don't think I can choose

572
00:54:23,760 --> 00:54:31,600
if the Microsoft come come to you and say you get all the money and resources you need which

573
00:54:31,600 --> 00:54:39,520
feature will you develop I think like it's probably will be in front of you because I'm using

574
00:54:39,520 --> 00:54:45,520
that you know I will probably expand like the memory management feature like a lot more into

575
00:54:45,520 --> 00:54:56,080
a lot more capabilities yeah yeah then I come to my clothing questions the question what is the

576
00:54:56,080 --> 00:55:04,800
yeah the the yeah the biggest misunderstanding people have about AI

577
00:55:07,680 --> 00:55:14,400
like yeah like that it can solve like it can solve many problems but like it's not like it can it

578
00:55:14,400 --> 00:55:21,280
can't it needs to be in every everything like it needs to you're around if I use case for it to

579
00:55:21,280 --> 00:55:26,080
solve that problem for you so it will only help you you know where to use it effectively and also

580
00:55:26,080 --> 00:55:32,800
in the way you want to use it effectively yeah awesome thank you also for joining me today I really

581
00:55:32,800 --> 00:55:39,680
enjoyed this conversation because it showed up that successfully I isn't just a large language model

582
00:55:39,680 --> 00:55:44,560
or latest announcement it's about combinating trust data responsible governance solid

583
00:55:44,560 --> 00:55:50,960
engineering and the right business mindset so yeah thank you for for being here and yeah all

584
00:55:50,960 --> 00:55:57,600
all the listeners you find all the links in the show notes so you can connect this to Paggy

585
00:55:58,720 --> 00:56:05,760
and yeah if you enjoyed this episode so please leave a review and tell us your your colleagues and

586
00:56:05,760 --> 00:56:11,440
yeah thank you for listening and thank you so I for having you here. Thank you for having me.

587
00:56:11,440 --> 00:56:17,200
Yeah he was great chatting with you. Yeah good bye bye.