Beyond Copilot: Building AI That Works in the Real World with Azure AI Foundry and Agentic AI for Frontline Workers with Fergus Kidd [MVP]

1
00:00:00,000 --> 00:00:05,800
Welcome back to the M665FM, where we sit down with Microsoft MWP's engineers, founders,

2
00:00:05,800 --> 00:00:10,560
and industry experts to explore the technology shaping the future of work.

3
00:00:10,560 --> 00:00:18,800
Today's guest is for this kit, Microsoft MWP co-founder and city of the Paul AI, and someone

4
00:00:18,800 --> 00:00:24,400
with a fascinating journey from development computer vision systems for the X or Mars mission

5
00:00:24,400 --> 00:00:29,200
to build AI solutions that empowered frontline workers.

6
00:00:29,200 --> 00:00:40,400
We discuss today, Azure AI Foundry, a genetic AI, multi-model AI, Microsoft co-pilot, computer

7
00:00:40,400 --> 00:00:45,000
vision field operation, and what really takes to build AI systems that deliver measurable

8
00:00:45,000 --> 00:00:48,720
business value outside the traditional office environment.

9
00:00:48,720 --> 00:00:55,840
If you're interested in Azure, Microsoft AI, and so on, say, "Welcome to our guest."

10
00:00:55,840 --> 00:01:01,600
Let's dive in. Welcome, Fergus. Thank you. Thanks for having me on the podcast, Marko.

11
00:01:01,600 --> 00:01:06,240
That's one thing. I think, "Goods, it's really cool. You started your career working on the

12
00:01:06,240 --> 00:01:12,480
X or Mars rover, how about?" Yeah, and it's still in the warehouse. It's still not on

13
00:01:12,480 --> 00:01:19,840
Mars, which is the most disappointing thing, I think. But yeah, when I was at UCL in London,

14
00:01:19,840 --> 00:01:25,440
I got the opportunity to work on some of the software for the camera system, which is kind

15
00:01:25,440 --> 00:01:33,600
of what triggered my journey into AI. That was about 2013, 2014, so it's pretty early

16
00:01:33,600 --> 00:01:38,320
on compared to what we have today, right, with large language, more agents everywhere.

17
00:01:38,320 --> 00:01:46,040
We were really looking at how, there's a substantial delay between communications from the

18
00:01:46,040 --> 00:01:54,840
Earth to Mars, so how can the rover itself do more at the edge of the frontline when it's

19
00:01:54,840 --> 00:02:01,240
driving around? Sadly, it's never left Earth because of various geopolitical things that

20
00:02:01,240 --> 00:02:05,880
have happened since then, but maybe one day, it's fairly sure they built it.

21
00:02:05,880 --> 00:02:24,040
For short, I read a joke, why astronauts, by rockets, run on Linux because they cannot open

22
00:02:24,040 --> 00:02:37,000
the window in the space. I think a master of them are not working on Windows or Thoms thing.

23
00:02:37,000 --> 00:02:43,560
How was your journey into the Microsoft tech ecosystem? Yeah, that's a great question. After that,

24
00:02:43,560 --> 00:02:49,560
I spent a couple of years working in the space and defense sectors, but mostly working on

25
00:02:49,560 --> 00:02:58,440
programmatic AI, so looking at machine learning models, Python. It was obvious very early on when

26
00:02:58,440 --> 00:03:04,840
you're working in that AI space that compute was the limiting factor, right? We think about

27
00:03:04,840 --> 00:03:11,240
the Mars rover. We had huge compute limitations because they could only send a certain size of hardware

28
00:03:11,240 --> 00:03:17,880
out there. As we started to look at other projects, especially around computer vision at that time,

29
00:03:18,440 --> 00:03:25,880
we had to look to cloud. Microsoft was an obvious choice. Azure was still young at that point,

30
00:03:25,880 --> 00:03:33,800
but have been involved in the early days. AWS was a big competitor, but I focused in on Azure.

31
00:03:33,800 --> 00:03:44,040
I guess the main one that you'll be familiar with is just everybody is there. Everybody is in

32
00:03:44,040 --> 00:03:50,280
Microsoft. Everybody has an outlook account or teams or something, so it's where everyone is,

33
00:03:50,280 --> 00:03:53,480
and it's where the helpful stuff is for doing our jobs, I guess.

34
00:03:53,480 --> 00:04:03,080
Yeah, this is awesome. So when you look back, what lesson from space research still influenced

35
00:04:03,080 --> 00:04:08,040
your work today? That's a really good question. I think definitely that idea of,

36
00:04:10,040 --> 00:04:17,640
well, I'd say two things. One is that the XMR's rover is just a good example of a worker on the

37
00:04:17,640 --> 00:04:21,800
front line, right? It's a pretty unique front line and it's a pretty remote environment.

38
00:04:21,800 --> 00:04:29,800
But the same things that we need to consider for a rover on a different planet,

39
00:04:29,800 --> 00:04:35,240
I'll kind of true for the kind of workers that I'm working with today. So we work a lot with

40
00:04:35,240 --> 00:04:41,400
garage technicians. We work a lot with health and safety people who might be upwind turbines

41
00:04:41,400 --> 00:04:47,720
in the North Sea, scaling electricity infrastructure at the top of Pylons.

42
00:04:47,720 --> 00:04:51,720
And actually, the problems are the same. There's a limited amount of stuff that they can carry with

43
00:04:51,720 --> 00:04:56,840
them. There's a limited amount of connectivity. They're not going to be connected to 24/7.

44
00:04:56,840 --> 00:05:04,040
Yet, they're expected to be doing more and more and more, making more repairs faster,

45
00:05:04,040 --> 00:05:11,480
documenting what they're doing, all sorts of stuff. So in a kind of weird way, they're very similar.

46
00:05:11,480 --> 00:05:17,640
But I think what really carries over in my interests in computing an AI is that ability to do more

47
00:05:17,640 --> 00:05:26,680
at the edge. So I think we've seen over the last year or so, particularly a really interesting

48
00:05:26,680 --> 00:05:35,400
shift back to on-premise, on edge, on device, for the things that can be done there, as you're in the

49
00:05:35,400 --> 00:05:41,560
cloud, amazing tools. But if we've got the ability to do stuff at the edge, we kind of should,

50
00:05:41,560 --> 00:05:50,680
because it's quicker, faster, cheaper, and more resilient in terms of your closer to the person

51
00:05:50,680 --> 00:05:53,720
you're trying to help. So if you can do something at the edge and not send it off to the cloud,

52
00:05:53,720 --> 00:06:00,760
that's a great opportunity. Yeah, you have also your background as physics. Does this shape your

53
00:06:00,760 --> 00:06:07,000
approach to engineering? I think so. I'm not sure if it's kind of worn off now after all these years,

54
00:06:07,000 --> 00:06:13,720
but I think so. I think it, what I always say to people joining the workforce or interested in

55
00:06:13,720 --> 00:06:20,920
computing in AI is that that kind of scientific grounding gives you a really analytical approach,

56
00:06:21,800 --> 00:06:29,000
a statistical approach. I think especially today, you probably see it with your guests all the time,

57
00:06:29,000 --> 00:06:33,720
there's a lot of people building things because it's cool. And I love building things because it's

58
00:06:33,720 --> 00:06:39,800
cool as part of why I'm an envy to eat right. But I think having that ability to kind of stay

59
00:06:39,800 --> 00:06:46,280
step back looks statistically at the outcomes that you're really trying to look at, especially in an

60
00:06:46,280 --> 00:06:52,600
enterprise situation. I think that's really helpful. So I'm really keen on like A/B testing when we

61
00:06:52,600 --> 00:06:58,920
change features in our platform. I'm really keen in like measuring real output results and generating

62
00:06:58,920 --> 00:07:05,480
like real statistics. I think that's definitely come from my background. Yeah, awesome. Yeah, yeah,

63
00:07:05,480 --> 00:07:13,400
you are a knowledge worker in most conversations around Microsoft AI focused on knowledge worker.

64
00:07:13,400 --> 00:07:17,080
Why do you believe frontline workers deserve more attention?

65
00:07:17,080 --> 00:07:25,560
Because there's more of them. It's just an obvious answer. I think, especially living and operating

66
00:07:25,560 --> 00:07:32,600
in the Microsoft ecosystem as we do, we kind of in this bubble of, oh well, everyone has teams,

67
00:07:32,600 --> 00:07:38,520
everyone has access to SharePoint, everybody uses Outlook. And that is true in our bubbles in

68
00:07:38,520 --> 00:07:46,120
our spheres. It was true for the last 10 years of work that I was doing at larger corporate firms.

69
00:07:46,120 --> 00:07:52,040
But it's not true for the majority of workers who are actually out there doing things. I mean,

70
00:07:52,040 --> 00:07:56,440
we have clients who don't even have email addresses, which blew my mind. I was like, how do you

71
00:07:56,440 --> 00:08:04,760
even send them a paycheck? But that's the reality of the world we live in. I can look up the numbers,

72
00:08:04,760 --> 00:08:11,720
but it's something like 80% of a workforce doesn't sit behind a desk. Yeah, the majority of what

73
00:08:11,720 --> 00:08:18,200
we often talk about in terms of AI and agents and co-pilot is focused at those people sat behind a desk.

74
00:08:18,200 --> 00:08:25,960
So I became really interested in how we can support those 80% of people, those frontline workers

75
00:08:25,960 --> 00:08:34,120
who are often in my experience over the last year in a bit, doing really, really inefficient things.

76
00:08:34,120 --> 00:08:39,800
Because they're not on devices all day, they're on computers all day. They're not using

77
00:08:39,800 --> 00:08:44,440
chat GPT in their spare time. They're doing stuff with their hands. They're fixing

78
00:08:44,440 --> 00:08:50,120
heavy machinery. They're doing inspections of industrial sites and things like that, where they

79
00:08:50,120 --> 00:08:55,400
just not able to carry around the laptop and have them with them all the time. So yeah, a lot of

80
00:08:55,400 --> 00:09:00,840
them are on phones and tablets and have access to some great stuff. But sometimes that's disconnected.

81
00:09:02,200 --> 00:09:08,280
Another thing that's come up in the last year and a half of me, which is super interesting is that

82
00:09:08,280 --> 00:09:17,480
when we're talking to managers and people who are sitting behind desks looking across a spectrum of work,

83
00:09:17,480 --> 00:09:23,480
and looking at what's going on, those people often have the answers that frontline workers are looking for.

84
00:09:23,480 --> 00:09:28,760
You know, frontline worker has a question. They probably today ring up a manager or try and find that

85
00:09:28,760 --> 00:09:35,640
information through a human. That person is sat at a desk and has access to it. And it's almost

86
00:09:35,640 --> 00:09:40,920
always in a share point repository somewhere. Right? So that information exists and it's

87
00:09:40,920 --> 00:09:47,080
searchable and it's findable, but it's not in the hands of the people necessarily who need it. So yeah,

88
00:09:47,080 --> 00:09:53,240
we feel how we kind of became really interested in this idea of like, how do we put the stuff that

89
00:09:53,240 --> 00:09:57,480
already exists, you know, the knowledge within organizations that already exists in the hands of

90
00:09:57,480 --> 00:10:06,840
our workers. Awesome. That's really, really interesting. What other do I say unique

91
00:10:06,840 --> 00:10:19,080
problems that frontline workers have and why traditional software often failed with,

92
00:10:19,880 --> 00:10:30,360
with field workers? Yeah, it's down to a number of reasons. I would say the number, there's a

93
00:10:30,360 --> 00:10:35,080
couple of big ones. One is connectivity. A lot of the people we're working with are only,

94
00:10:35,080 --> 00:10:42,440
you know, even in the world of 5G and mobile internet everywhere, they're only probably connected

95
00:10:42,440 --> 00:10:49,480
for like an hour out of an eight hour day, whether that was of actual mobile network or Wi-Fi

96
00:10:49,480 --> 00:10:53,160
coverage or whatever it is, but really it's because they're doing something else with their hands.

97
00:10:53,160 --> 00:10:58,440
They can't be at a device for these other hours a day, yet they're expected to do like a lot of

98
00:10:58,440 --> 00:11:08,440
reporting and other things. So it's like, well, how do we help them if they're not that kind of

99
00:11:08,440 --> 00:11:13,640
connected? There's a few ways we can do that. So we work a lot with realware who produce these kind

100
00:11:13,640 --> 00:11:19,320
of headsets, like wearable compute headsets. So it's kind of like having a tablet on your

101
00:11:19,320 --> 00:11:25,480
head at all times using that voice, you know, voice first AI systems and things like that to actually

102
00:11:25,480 --> 00:11:33,560
actually help here. Awesome, that's really cool. That is really cool.

103
00:11:33,560 --> 00:11:40,360
Yeah, they're super useful as well because they, you know, they completely voice controlled,

104
00:11:40,360 --> 00:11:44,920
so I can be doing stuff with my hand. I've got a camera, even a thermal camera up here.

105
00:11:44,920 --> 00:11:49,400
And there's also different versions of wearable. This is a device from a company called Music,

106
00:11:49,400 --> 00:11:52,920
which is obviously a lot smaller and just kind of sits here, kind of,

107
00:11:52,920 --> 00:11:59,880
char hard hat or whatever, PPE. And then the direct, you get the information direct in your

108
00:11:59,880 --> 00:12:05,720
your eyes or how is it working? If you're connected, right? So if you're connected to, I mean, there's

109
00:12:05,720 --> 00:12:09,720
a number of things we can do on here. So like, realware have teams integration, for example, they have an app.

110
00:12:09,720 --> 00:12:15,720
I can just say call my manager and like, I'll pop into them on a team school, which is really great.

111
00:12:15,720 --> 00:12:21,560
And a lot, a lot of people are using that already. But yeah, if I'm disconnected, how do I answer

112
00:12:21,560 --> 00:12:27,640
questions is it's kind of like our biggest challenge. But the second biggest challenge with

113
00:12:27,640 --> 00:12:31,960
frontline workers is because they're not using, they're not sat at a computer, they're not using

114
00:12:31,960 --> 00:12:37,000
mobile devices all day every day. A lot of them are kind of unfamiliar with using devices like these.

115
00:12:37,000 --> 00:12:44,280
They're used to using pen and paper. So they kind of all default to that. So our big focus is just

116
00:12:44,280 --> 00:12:49,800
make everything as easy as possible. If we have an app on here, make it as easy as possible to use.

117
00:12:49,800 --> 00:12:53,480
If we have an app on the phones or the tablets, make it as easy as possible to use.

118
00:12:53,480 --> 00:13:01,720
Like, yeah, that's a huge barrier to entry is I have to learn another tool. And what I found really

119
00:13:01,720 --> 00:13:05,720
interesting is like working with people in automotive, you give them a tablet and they're like,

120
00:13:06,600 --> 00:13:11,640
need to learn, I need to learn how to use it. I need to invest time and energy into using it.

121
00:13:11,640 --> 00:13:16,520
But show another a tablet that gets like car diagnostic readings out or some other like highly

122
00:13:16,520 --> 00:13:22,040
technical information. They pick it up like that. So for me, for me, it's kind of like, how do we

123
00:13:22,040 --> 00:13:27,880
reframe this is not a complicated system, but just a tool that you're using to do a job, right?

124
00:13:27,880 --> 00:13:34,920
I think that's quite important as well. Yeah, I think also Microsoft had built such a tool. It's,

125
00:13:34,920 --> 00:13:38,440
I think it's called HoloLens, I don't know what's happened there.

126
00:13:38,440 --> 00:13:44,840
Well, you'll have to get somebody more than know than me on that. But yeah, I mean,

127
00:13:44,840 --> 00:13:49,240
what five years ago now, maybe we were doing a lot of work on HoloLens as a fully kind of

128
00:13:49,240 --> 00:13:55,240
AR spatial platform. And it had some really, really interesting use cases.

129
00:13:55,240 --> 00:14:02,600
And obviously like full spatial awareness. So if anybody listening or watching

130
00:14:03,240 --> 00:14:07,480
hasn't tried one on, there's basically like a full headset, Google with a

131
00:14:07,480 --> 00:14:13,080
through display, but you can, you can get 3D spatial awareness. Really amazing doing some of the

132
00:14:13,080 --> 00:14:17,880
stuff that we're doing at Phil Powell, which is around inspecting site inspections and things.

133
00:14:17,880 --> 00:14:25,800
But for the price and the learning curve, a lot of people, a lot of technicians in those

134
00:14:25,800 --> 00:14:31,720
situations just prefer to use mobile devices, right? I can speak into it. I can take a photo. I can

135
00:14:31,720 --> 00:14:38,280
write any notes that I have. I'm kind of done. The HoloLens is a full AR platform was quite heavy.

136
00:14:38,280 --> 00:14:43,880
The most thing about these, I would describe these more as like wearable compute than

137
00:14:43,880 --> 00:14:49,080
augmented reality. They just sit in front of you and give you access to the information

138
00:14:49,080 --> 00:14:53,400
when you're looking down at it. So you can ignore, I could talk into you, I can ignore it completely

139
00:14:53,400 --> 00:14:57,240
or I can kind of glance down and see a small screen, there's some information on it.

140
00:14:58,280 --> 00:15:08,120
For us, the real unlock of potential has been about voice. Everybody's used to using their voice

141
00:15:08,120 --> 00:15:14,440
to talk to other humans. So allowing us to talk to, for things that we need to get done, whether

142
00:15:14,440 --> 00:15:18,680
that's looking up information or whether that's reporting on information, I think that's,

143
00:15:18,680 --> 00:15:25,000
you know, the HoloLens is a great device, but it was kind of overkill for some of the things that

144
00:15:25,000 --> 00:15:31,960
we're trying to do. Did you know if they still develop the HoloLens or is it also retried?

145
00:15:31,960 --> 00:15:40,040
As far as I'm aware, it's retired. I think there's still use cases in which HoloLens 2 are still

146
00:15:40,040 --> 00:15:47,880
operating. I know they did. I did some really interesting mentoring for a project at UCL,

147
00:15:47,880 --> 00:15:53,880
where students were putting training videos and like 3D models in HoloLens's for the medical

148
00:15:53,880 --> 00:15:58,840
sector. So people could kind of get an idea of they're in the middle of a surgery, like what that

149
00:15:58,840 --> 00:16:03,400
looks like and have a much more immersive experience. There's definitely still products going on,

150
00:16:03,400 --> 00:16:08,120
like, on like that, but as for the production of like a HoloLens 3, I don't think we'll see that

151
00:16:08,120 --> 00:16:15,800
sadly. But I've, you know, we work with a lot of wearable compute partners outside of Microsoft,

152
00:16:15,800 --> 00:16:20,840
and I think there are some really exciting things going on in the hardware space. I think it won't

153
00:16:20,840 --> 00:16:26,200
be too long before we see kind of like products like the meta glasses come along with like screens

154
00:16:26,200 --> 00:16:30,760
and all sorts of cool stuff in that format, which is really, really interesting as well.

155
00:16:30,760 --> 00:16:42,200
Yeah, that's cool. I think the meta glat, the what was it because Google starts really early,

156
00:16:42,200 --> 00:16:49,160
was building this. I don't know how I forgot the name from the smart. Yeah, it was just a little

157
00:16:49,160 --> 00:17:02,360
thing. Yeah. And I think also Facebook or meta has the meta quest, I think. And yeah, there

158
00:17:02,360 --> 00:17:11,960
and there it's looked like there's upcoming beside these big players, all these big players.

159
00:17:11,960 --> 00:17:18,360
We try out the most project, I think. A lot of smart companies coming up with smart,

160
00:17:18,360 --> 00:17:24,680
glasses. What was, why did you think the big ones failed so heavily?

161
00:17:24,680 --> 00:17:33,480
I think, honestly, I think the HoloLens is separate. I think that comes down to the cost

162
00:17:33,480 --> 00:17:38,280
manufacturer. Those were really high-end innovation devices, and they were just expensive for

163
00:17:38,280 --> 00:17:45,880
the enterprise. You know, what we're finding is customers of hours who have technicians in the

164
00:17:45,880 --> 00:17:50,520
field, do not want to spend any money on devices or hardware because they will get broken.

165
00:17:50,520 --> 00:17:56,040
The real, we work with real-wearers because these are fully ruggedized, like you can smash these

166
00:17:56,040 --> 00:18:02,440
to pieces and they will carry on working, hopefully. They're really, and there's even a more

167
00:18:02,440 --> 00:18:09,080
ruggedized version, and they've got all sorts of accreditations, etc. So that's the only one I've

168
00:18:09,080 --> 00:18:15,160
conceived, like a really working in those heavy industries, just because of the form factor and

169
00:18:15,880 --> 00:18:20,680
the kind of damage control. A HoloLens $3,000, I wouldn't want to drop it on the floor,

170
00:18:20,680 --> 00:18:26,200
you know, but that's what happens in the reality for these devices. Why I see the other big players,

171
00:18:26,200 --> 00:18:32,440
like Meta and the Google Glass failing, is they just not write the right devices for the enterprise,

172
00:18:32,440 --> 00:18:38,040
right? As a developer, I can't easily build enterprise apps for them. They don't easily integrate

173
00:18:38,040 --> 00:18:45,080
into where enterprise workers are. So I can't, for example, very easily connect my Meta,

174
00:18:45,720 --> 00:18:49,720
glasses through an app to dump photos into a SharePoint repository.

175
00:18:49,720 --> 00:18:54,920
Meta, for it's not useful to the enterprise, right? Whereas on the real-wear and the

176
00:18:54,920 --> 00:19:00,360
music and other companies that are producing those, I'm free to build my own Android,

177
00:19:00,360 --> 00:19:05,640
they're all Android devices, so I'm free to build my own apps, and then I can use the integrations

178
00:19:05,640 --> 00:19:12,040
through Azure or through the Microsoft ecosystem to actually build kind of like a custom app that

179
00:19:12,040 --> 00:19:17,080
works for me in my enterprise use case. Because otherwise, they're just off the shelf consumer devices.

180
00:19:17,080 --> 00:19:25,000
You know, we get, we get asked questions a lot for reporting. Okay, so say I'm at a chemical

181
00:19:25,000 --> 00:19:29,240
manufacturing site, and I want to go around, and I need to take photos of certain pieces of machinery

182
00:19:29,240 --> 00:19:33,400
and provide fixes. People say, "Oh, couldn't I wear a pair of Meta glasses, and that just takes

183
00:19:33,400 --> 00:19:39,240
a photo of what I'm looking at?" Yes, but I can't integrate that into my, into our app, or your back-end

184
00:19:39,240 --> 00:19:44,920
systems. So, what do you want to do? Take a hundred photos and drag and drop them, upload them to Google

185
00:19:44,920 --> 00:19:50,280
Drive, and download them from Google Drive. It just, the flow is interrupted, and then it never takes

186
00:19:50,280 --> 00:19:57,800
off beyond like a roof of content. Yeah, but for the entry resource, I think there was, I don't know,

187
00:19:57,800 --> 00:20:03,800
Meta has invests, I think, billions of dollars in there. We are environment,

188
00:20:04,520 --> 00:20:15,160
Bicolceft tells, "Build mesh." Okay, it's now, no, in teams, the function, why did they fail so hard?

189
00:20:15,160 --> 00:20:25,880
I think it's just wrong, wrong product for the audiences. I do believe that Meta will continue to make,

190
00:20:25,880 --> 00:20:31,160
you know, our gaming platform. I think that's a great use of the technology. It's just a consumer market.

191
00:20:34,440 --> 00:20:41,480
Do people really want meetings in 3D virtual spaces? You know, we did it quite a lot when I was

192
00:20:41,480 --> 00:20:45,880
working at Avernard. It was a bit gimmicky, and then nobody ever scheduled it that way.

193
00:20:45,880 --> 00:20:52,360
Going forward, you know? It was a great toy to play with. And then the second reason is just the cost,

194
00:20:52,360 --> 00:20:57,720
you know, that was a premium feature. Teams, to do any customization work, you needed to have

195
00:20:57,720 --> 00:21:02,920
special add-ons in Azure, which we didn't build at the time. And there were expensive to run,

196
00:21:02,920 --> 00:21:07,640
and the 3D environments were kind of expensive to run. So versus just chatting on a web call,

197
00:21:07,640 --> 00:21:12,600
is the benefit there? No, not really. And it's the same reason why people aren't

198
00:21:12,600 --> 00:21:18,360
realistically using HoloLens devices in industrial settings, or aren't using MetaQuests

199
00:21:18,360 --> 00:21:23,000
outside of training rooms and training environments, just because it's just not the right solution

200
00:21:23,000 --> 00:21:30,760
for the enterprise, Adley. And yeah, that is really cool. Let's talk a little bit about a field

201
00:21:30,760 --> 00:21:39,800
poll AI. What inspired you to start field poll AI, and which customer problems did you try to solve first?

202
00:21:39,800 --> 00:21:47,640
Yeah, so our founder James actually came out of realware. He was head of software there, and so he

203
00:21:47,640 --> 00:21:54,120
had a lot of touch points with clients, all asking the same two questions, which is, I want to look

204
00:21:54,120 --> 00:22:00,440
up data, and I want to fill in forms in the field in an easier way. That's just done badly across

205
00:22:00,440 --> 00:22:04,840
the board. A lot of the times, lots of clients were working with it, still done on paper and pen.

206
00:22:04,840 --> 00:22:11,000
And he just had this idea, well, wouldn't, isn't AI perfect for that?

207
00:22:11,000 --> 00:22:15,800
In a world where I could just speak into a device, and it could understand,

208
00:22:15,800 --> 00:22:21,960
and I'll put quotation marks there, and understand, it could understand what's going on in order

209
00:22:21,960 --> 00:22:27,160
to help me do the jobs and tasks that I need to do quicker, whether that's looking up information

210
00:22:27,160 --> 00:22:33,560
or reporting on something. And so we kind of just went from there, and I think the most exciting thing

211
00:22:33,560 --> 00:22:38,680
has been going out to real front-line workers and showing it to them. Their kind of initial reaction

212
00:22:38,680 --> 00:22:44,520
is always, oh, another, another tool, another system, another, they then actually see it working,

213
00:22:44,520 --> 00:22:49,880
and they love it, because instead of having to sit there for 20 minutes and write something

214
00:22:49,880 --> 00:22:54,680
up, they can just have a three-minute conversation, and all of that reporting and write-up is done for

215
00:22:54,680 --> 00:23:00,280
them. The managers in front of how stuff love it, because they're no longer having to read and

216
00:23:00,280 --> 00:23:08,520
retip handwritten forms. We had one client who do video inspection surveys, and some of their front

217
00:23:08,520 --> 00:23:14,440
of how stuff, we're watching those videos four times just to type out the key information that's in

218
00:23:14,440 --> 00:23:20,120
them, and we said, no, just transcribe the video, and then we'll create a summary for you.

219
00:23:20,120 --> 00:23:25,400
You know, it's quick, super easy, and then that frees them up to do more important work,

220
00:23:25,400 --> 00:23:31,560
like calling their clients and asking questions. It gives them much more time for that human

221
00:23:31,560 --> 00:23:36,200
interaction, and the text in the workshop, it gives them more time to get on with doing what

222
00:23:36,200 --> 00:23:43,080
they're good at, which is fixing stuff or teaching apprentices or whatever it is that they could spend

223
00:23:43,080 --> 00:23:49,960
more time doing than paperwork. The other thing is that it's just such a broad opportunity,

224
00:23:49,960 --> 00:23:56,840
because anybody who works for a company and is not sat behind a computer is definitely doing

225
00:23:56,840 --> 00:24:01,800
paperwork or looking up information on how to fix things, how to do things, and just think of the

226
00:24:01,800 --> 00:24:07,720
inefficiencies in your own life. I mean, we've all got a story where, you know, I don't know, a fridge

227
00:24:07,720 --> 00:24:12,600
broke down, and somebody came out to repair it and said, I don't have the correct part, because I

228
00:24:12,600 --> 00:24:17,320
didn't know what the model number was. They come back two weeks later, and they realized that part

229
00:24:17,320 --> 00:24:23,960
doesn't fit in that particular, you know, there's so many inefficiencies in the world around us that

230
00:24:23,960 --> 00:24:31,960
are ripe for fixing and sorting out, and AI, I believe, is a huge driver of finding and fixing,

231
00:24:31,960 --> 00:24:37,880
because it inefficiencies, because one question we have to ask ourselves is, well, people have had

232
00:24:37,880 --> 00:24:45,240
tablets for decades now. Why are we still paying things in paper and pen? And the answer is,

233
00:24:45,800 --> 00:24:51,240
because it's faster than typing away on a tablet if you're not used to typing, what we found is

234
00:24:51,240 --> 00:24:58,200
everybody's used to talking, we do it all the time. So add in a bit of AI and you take out that

235
00:24:58,200 --> 00:25:03,320
middle step and you can make genuine efficiencies savings and make and streamline your business. So,

236
00:25:03,320 --> 00:25:09,480
yes, it's been super rewarding to kind of see, people actually use the product and actually love it.

237
00:25:11,320 --> 00:25:18,680
I think a little bit about it is the future that the people run to the streets and talk to themselves.

238
00:25:18,680 --> 00:25:25,320
Well, we said the same thing in the early 2000s when people started wearing, you know,

239
00:25:25,320 --> 00:25:29,000
Bluetooth headsets and running down the streets. There's always that moment of like a big focus in

240
00:25:29,000 --> 00:25:37,160
me or a big focus for themselves. But, you know, for us, I'll give you a real-world example. We're working with

241
00:25:37,160 --> 00:25:44,280
a company that installs doors and windows and they do that all over the country, they're UK-based,

242
00:25:44,280 --> 00:25:50,200
they do it all over the country. But for every single installation, they have to do some paperwork

243
00:25:50,200 --> 00:25:55,240
for themselves internally to know that what was put where, what serial number of windows was installed,

244
00:25:55,240 --> 00:26:01,240
what type of glass they put in. But also for the government, you know, for the grants that they

245
00:26:01,240 --> 00:26:07,960
give out for making homes more efficient, the government has forms that they need to do. So,

246
00:26:07,960 --> 00:26:14,200
what they were doing was four days a week and installer would be out on the road driving to

247
00:26:14,200 --> 00:26:20,200
people's homes, completing jobs, taking measurements, going back to the workshop, completing jobs.

248
00:26:20,200 --> 00:26:26,440
And then on the fifth day, they would sit in a hotel room and take all of the photos they'd

249
00:26:26,440 --> 00:26:31,400
taken off their phones, go back to, so say they're doing that on Friday, they'd go back to Monday and

250
00:26:31,400 --> 00:26:36,520
write up a report of what they did, then they go to Tuesday, then, and that would take a full day.

251
00:26:36,520 --> 00:26:43,000
Not only at inefficient because you're spending a fifth of your time doing paperwork when your job is

252
00:26:43,000 --> 00:26:49,160
a stroller of doors and windows. But also, you're going to forget information. There's going to be

253
00:26:49,160 --> 00:26:53,560
by Friday, you're going to forget what actually happened on Monday or why you put this

254
00:26:53,560 --> 00:27:00,120
that in that window in backwards or whatever, because they were real physical constraints,

255
00:27:00,120 --> 00:27:04,440
you're just not going to remember. So, our product, because we have a conversation with them

256
00:27:04,440 --> 00:27:10,920
through the AI, when they're on site at the job, we can collect all of the information without

257
00:27:10,920 --> 00:27:16,120
having to spend another day on necessarily filling in the paperwork, but we can capture all of the

258
00:27:16,120 --> 00:27:21,080
information. And the AI can actually guide them, right? So, it's like you haven't taken a photo of

259
00:27:21,080 --> 00:27:25,080
the final install, you haven't given me the serial number. If there's a missing information,

260
00:27:25,080 --> 00:27:31,080
that's caught at the point of entry for that data. So, it's not just speeding things up, but it's

261
00:27:31,080 --> 00:27:36,680
actually also improving data quality. So, by the time it does get to us behind the desk as knowledge

262
00:27:36,680 --> 00:27:47,320
workers, it's actually there in a good format and it's clean. Awesome. I think another hot topic,

263
00:27:47,320 --> 00:27:52,520
actually, it's all talked about as a Genetic AI, how does it fit into your platform?

264
00:27:52,520 --> 00:27:57,720
Yeah, I think to be honest, a Genetic AI has really been the catalyst in allowing a product

265
00:27:57,720 --> 00:28:06,760
like fieldpal to exist, because it really elevates an AI product from a chatbot

266
00:28:06,760 --> 00:28:13,720
into a usable system or platform, which is what we have, right? So, our platform has multiple

267
00:28:13,720 --> 00:28:19,960
capabilities. We can look up information from knowledge bases, API anywhere a customer has data,

268
00:28:19,960 --> 00:28:24,280
we can look it up, whether that's from like a SharePoint repository, whether that's from a life, API,

269
00:28:24,280 --> 00:28:30,600
we can also help people take notes and things and then also fill in reports.

270
00:28:30,600 --> 00:28:35,560
Those are fundamentally for an AI, there's a fundamentally very different things. The way that we

271
00:28:35,560 --> 00:28:41,080
want the AI to respond and talk to our users is completely different depending on the task that

272
00:28:41,080 --> 00:28:46,920
they're doing. One thing is making it agentic, we have those individual agents that present

273
00:28:46,920 --> 00:28:53,400
themselves through an orchestrator as one thing. So, if I say, "I've got a broken down fridge,

274
00:28:53,400 --> 00:28:59,800
it's the power supply unit, what do I do?" It will know which agent has what capabilities and

275
00:28:59,800 --> 00:29:06,200
where to look for that data without opening up a whole kind of worms of, "Let's try and do

276
00:29:06,200 --> 00:29:15,000
XYZ at the same time." So, we can really make bespoke products per client because we can have agents

277
00:29:15,000 --> 00:29:21,160
that only exist in certain clients' environments. So, if I'm an automated garage, I can have a system

278
00:29:21,160 --> 00:29:28,520
that is able to look up part numbers and prices. If I'm doing health and safety inspections on

279
00:29:28,520 --> 00:29:33,720
a wind farm, that's going to be a completely different agent because it's got a whole host of

280
00:29:33,720 --> 00:29:39,720
different abilities. But to present themselves as one friendly chatbot, I hate the word chatbot

281
00:29:39,720 --> 00:29:45,480
now because of agent to go out, but to present itself as one interface, if you like, and then we can

282
00:29:45,480 --> 00:29:53,320
turn off and on agents. And then the other thing is the ability to preempt, use his actions, and do things

283
00:29:53,320 --> 00:29:59,480
before they turn up to a job site and have those agents run and doing things in the background as a

284
00:29:59,480 --> 00:30:06,440
game changer as well. And it really just, you know, we've all been frustrated on those websites where

285
00:30:06,440 --> 00:30:12,760
they've got a chatbot and you can kind of click one or two replies. Genetic AI for us has changed the

286
00:30:12,760 --> 00:30:18,280
game into a purely conversational, you can talk back or forth, it will tell you what it can and can't do,

287
00:30:18,280 --> 00:30:24,120
it will know what tools it can call and what APIs and data it has access to in a really kind of

288
00:30:24,120 --> 00:30:34,440
flexible and scalable way. So it's, yeah, a complete game changer. Yeah. So you are an MVP in AI. So I

289
00:30:34,440 --> 00:30:43,160
think you're, you get it when you are using Azure AI Foundry. Have you built up the company with Azure AI Foundry?

290
00:30:43,160 --> 00:30:51,480
Yeah. So all of our models are in Azure AI Foundry. Our whole product sits on top of Azure. So it's

291
00:30:51,480 --> 00:30:58,280
there's architecturally, there's a lot of different bits and pieces now. For example, we,

292
00:30:58,280 --> 00:31:03,800
so we do a process that we call microindexing, which is kind of like a secret source on top of some

293
00:31:03,800 --> 00:31:11,160
of the retrieval augmented generation stuff that we're doing within Azure. So things like cognitive

294
00:31:11,160 --> 00:31:15,880
search, we do some pre-processing steps, but those are actually done by AI agents,

295
00:31:16,760 --> 00:31:21,880
completely separately in the platform and those all sit in Azure Foundry and talk to each other.

296
00:31:21,880 --> 00:31:30,040
And then obviously a huge part of our platform is the speech, which is all through Azure AI

297
00:31:30,040 --> 00:31:34,920
Services and Azure AI Foundry as well. Then we've got lots of boring stuff like

298
00:31:34,920 --> 00:31:45,480
Kubernetes clusters and web apps and etc, etc, etc, etc, to fund it, like to host the actual platform

299
00:31:45,480 --> 00:31:51,160
itself as well. But yeah, we're partnered with Microsoft. Microsoft have been really, really supportive.

300
00:31:51,160 --> 00:31:55,560
So if there's anybody listening or watching that's thinking, well, I have a really good idea

301
00:31:55,560 --> 00:32:01,960
for a company that's got AI, it's called Microsoft has some really great supported programs for

302
00:32:01,960 --> 00:32:06,200
helping you through that, giving you some Azure credit, giving you some kind of help and advice.

303
00:32:06,200 --> 00:32:10,760
So I definitely encourage you to look into that. I mean, other cloud platforms are available,

304
00:32:10,760 --> 00:32:16,360
but I really know about the process. With Microsoft. This is still a startup program, all right?

305
00:32:16,360 --> 00:32:23,320
Yes, yes. There's a number of different startup programs that they have, some really early ones,

306
00:32:23,320 --> 00:32:28,680
where you can get $5,000 of Azure credit just to kind of pad out an idea and give it a real architecture.

307
00:32:28,680 --> 00:32:34,360
And then there's also some industry back startup programs, so we're lucky enough to be backed by

308
00:32:34,360 --> 00:32:40,680
Barclays, which is a bank in the UK through Microsoft for some more Azure credit from things.

309
00:32:40,680 --> 00:32:48,600
And then through the NDP network, where you can go and find an NDP and find one for yourself,

310
00:32:48,600 --> 00:32:54,760
but through the NDP network, we've got access to an amazing set of people like ourselves

311
00:32:54,760 --> 00:32:58,760
who have the knowledge and the skills and are kind of happy to share it.

312
00:32:59,960 --> 00:33:09,880
Yeah, what I have seen is also some companies that build AI products, they're taking them to the show

313
00:33:09,880 --> 00:33:14,280
and promote it also, they're doing also marketing for those. It's really amazing.

314
00:33:14,280 --> 00:33:21,320
So, yeah, we're lucky enough to be in that position. So we are Microsoft's demo of choice in the

315
00:33:21,320 --> 00:33:28,680
Microsoft Innovation Hub for industrial workers. So if you do find yourself in the Microsoft

316
00:33:28,680 --> 00:33:35,000
Innovation Hub, you can ask for a demo or feelpile on the tablets. And that's amazing. That's opening us up to

317
00:33:35,000 --> 00:33:43,160
lots of different clients in lots of different sectors. But it's because we've made a very

318
00:33:43,160 --> 00:33:47,960
kind of usable user-friendly demo that makes sense to a lot of people. Oh, I can fill in the form

319
00:33:47,960 --> 00:33:53,000
really, really quickly just using my voice. It's a really nice thing to demo, so Microsoft

320
00:33:53,000 --> 00:34:00,120
being kind of left to be doing that across the world for us here. Yeah, I think Microsoft had also

321
00:34:00,120 --> 00:34:11,080
what's called co-pilot studio to building the AI agents and they have Foundry. I did

322
00:34:11,080 --> 00:34:17,480
choose Azure AI Foundry. What's the reason and why is it better for you?

323
00:34:18,360 --> 00:34:24,040
Really, really good question. I've promised I didn't write these because that's exactly why I would have

324
00:34:24,040 --> 00:34:34,360
asked myself. So for us, when we are working with workers and individuals who aren't even using

325
00:34:34,360 --> 00:34:38,760
teens regularly, which is kind of ape into you and I because we're on teams all the time,

326
00:34:38,760 --> 00:34:45,240
pinging each other, but you know, that's our kind of world now. But when you're a car mechanic or

327
00:34:45,240 --> 00:34:49,080
and you're just not using teams, if you have a question, you probably just go into another room or

328
00:34:49,080 --> 00:34:56,360
you would call someone up on the mobile, you know, they're just not used to that platform. And

329
00:34:56,360 --> 00:35:02,840
we have the thumb to them at AR. So whether they're using the headsets, whether they're using tablets,

330
00:35:02,840 --> 00:35:09,880
we needed to make a kind of fully end-to-end platform where we can do everything we need to do,

331
00:35:09,880 --> 00:35:16,200
including recording videos whilst talking to R.A.I. Uploading offline videos to be transcribed later,

332
00:35:16,200 --> 00:35:25,960
like lots of different niche things. There are industries that I would actually go back to using

333
00:35:25,960 --> 00:35:34,680
Copilot Studio. Basically, anywhere where that front-line worker is using teams or M365 as their primary

334
00:35:34,680 --> 00:35:41,640
method of communication, actually having our product as in Copilot Studio and just making it

335
00:35:41,640 --> 00:35:47,880
available to them would be really powerful and beneficial. The sort of industries that I'm thinking

336
00:35:47,880 --> 00:35:56,440
of when I say that is like in-home care workers, they're using teams on their mobile, all the time to

337
00:35:56,440 --> 00:36:04,120
talk to senior members of staff, etc, to get their scheduling and rotors and going back and forwards.

338
00:36:04,120 --> 00:36:11,080
So that's a perfect opportunity. The type of front-line workers that we have focused on so far

339
00:36:11,080 --> 00:36:17,160
are not using teams every day, so we decided to build a kind of separate platform.

340
00:36:17,160 --> 00:36:22,360
But the nice thing about Azure AI Foundry and Copilot Studio is that should we ever need to build

341
00:36:22,360 --> 00:36:27,880
an agent that sits in Copilot, that's actually very easy to do now, because we've got all of our

342
00:36:27,880 --> 00:36:33,640
APIs and our AI and we can actually, through Copilot Studio, just link up that Copilot to our own

343
00:36:33,640 --> 00:36:39,560
back end AI. So if I ask a question through my tablet with the product, it's well not actually make

344
00:36:39,560 --> 00:36:46,760
a difference if I asked it through a team's agent, which is really nice because it's like we've decided

345
00:36:46,760 --> 00:36:52,520
to go one direction, but if we ever wanted to change that decision, it would be quite easy to do so.

346
00:36:52,520 --> 00:36:57,320
Thanks to the tools that we've got and the back ends, the systems that we've got and the integrations

347
00:36:57,320 --> 00:37:05,800
into the platform, that we've got really interesting, kind of even further down the line, is that we have

348
00:37:05,800 --> 00:37:14,200
some clients in use case areas where technicians don't even want an app. So we're actually thinking of,

349
00:37:14,200 --> 00:37:18,840
well, how can we turn our AI into like a phone agent that just calls you and says,

350
00:37:18,840 --> 00:37:23,960
tell me the information that I need to know because they don't even want to open the app and use it,

351
00:37:23,960 --> 00:37:29,000
they're so used to using their hands and doing it that way. But again, everybody's used to,

352
00:37:29,000 --> 00:37:34,280
in those industries is used to taking and making phone calls. So we're really interested to get our AI

353
00:37:34,280 --> 00:37:43,240
working through just bringing people up basically. I think the most people around the world using

354
00:37:43,240 --> 00:37:51,160
the large language models, what's your view on the small or special language models?

355
00:37:51,880 --> 00:37:58,200
Yeah, so the core of our product is actually a small language model. We do not use our language model.

356
00:37:58,200 --> 00:38:08,920
Why? Because for us, domain expertise is incredibly important. We do not want, I'll give you

357
00:38:08,920 --> 00:38:13,240
an a real example from one of our clients. They said before finding us, they were playing around with

358
00:38:13,240 --> 00:38:18,600
ChatGbT, Alarm, already going off, but they're playing around with ChatGbT to answer some of their

359
00:38:18,600 --> 00:38:26,280
questions. And he said to ChatGbT, "Okay, I've got a land Rover in and it's air conditioning system

360
00:38:26,280 --> 00:38:33,320
isn't working. What do I do?" And ChatGbT produced, or I would describe it as a vomited out,

361
00:38:33,320 --> 00:38:39,640
a load of information not about land Rovers, about air conditioning systems in buildings and air

362
00:38:39,640 --> 00:38:46,760
conditioning in other cars. So that's actually less helpful than a system that just says, "I don't

363
00:38:46,760 --> 00:38:52,120
know." So the beautiful thing about small language models is that they don't always have this kind of

364
00:38:52,120 --> 00:38:59,240
inherent confidence that they know everything because their training sets are much smaller.

365
00:38:59,240 --> 00:39:08,520
And then we only expose what is available from our data sources. So we're only showing the AI,

366
00:39:08,520 --> 00:39:15,640
the small language models, real relevant data, that we, okay, in that case, we found it using

367
00:39:15,640 --> 00:39:19,880
Azure AI Search, which is a really, really great product, by the way. So if you're interested in

368
00:39:19,880 --> 00:39:27,640
search and semantic search, an amazing product, I love it. It will return really great relevant answers,

369
00:39:27,640 --> 00:39:32,440
and then all the small language model that is doing is kind of reading those answers and presenting

370
00:39:32,440 --> 00:39:39,240
the most relevant bits of information, but not in an imagined made up way. So for us, small language

371
00:39:39,240 --> 00:39:43,400
models are actually more powerful than large language models because they don't get distracted,

372
00:39:43,400 --> 00:39:49,560
they don't go off down on tangents, they only, they're much more focused in on what we want them to be

373
00:39:49,560 --> 00:39:54,200
focused in on. Now, the caveat to that is there's a lot more work to get all of that information in the

374
00:39:54,200 --> 00:40:01,080
right place. But once you do that, it's a lot more beneficial than just asking chat TBT or Gemini

375
00:40:01,080 --> 00:40:09,800
or whatever it is you're going to ask. Yeah, I think a little bit about, then you have for every client

376
00:40:09,800 --> 00:40:16,600
you have built a small language model because it's different to, I don't know, doors of windows and

377
00:40:16,600 --> 00:40:23,640
the other repair, I don't know, the refrigerers. So, so you have developed and for every client

378
00:40:23,640 --> 00:40:29,880
and own model. So today, we actually have just a core central language model that we kind of tweak

379
00:40:29,880 --> 00:40:35,880
and fine tune into those different industry use cases through some inventive prompting. So

380
00:40:36,520 --> 00:40:44,200
every single client runs off the same central model in an, an injury, but it presents itself in a

381
00:40:44,200 --> 00:40:48,920
slightly different way. So if I was from an automotive garage, it would know that it knew about

382
00:40:48,920 --> 00:40:56,360
cars and automotive versus if it was for like an appliance manufacturer, we would, we would know that.

383
00:40:56,360 --> 00:41:03,080
Our goal actually is to do what you're suggesting in that create kind of like domain specific

384
00:41:03,080 --> 00:41:10,120
knowledge, knowledgeable small language models that we can kind of drop in and drop out. Why? Because

385
00:41:10,120 --> 00:41:16,280
today we have this onus on our client that they are bringing all of their information. So if you

386
00:41:16,280 --> 00:41:21,080
want it to talk about land Rover air conditioning systems, you have to have access to all of that

387
00:41:21,080 --> 00:41:28,440
documentation. That isn't necessarily realistic. So there are some circumstances. The medical

388
00:41:28,440 --> 00:41:33,480
feels a great one Microsoft's done a load of work in creating like fine tuned specific language

389
00:41:33,480 --> 00:41:38,200
models that understand medical terminology. There are examples like that where we would actually want

390
00:41:38,200 --> 00:41:43,800
to swap out the small language models. But it makes the platform more complicated and difficult to

391
00:41:43,800 --> 00:41:49,720
administer, but it's definitely something we'll do in the future. The other benefit of a small

392
00:41:49,720 --> 00:41:54,440
language model is that one day we'll probably will have phones that are powerful to run versions

393
00:41:54,440 --> 00:41:59,000
of them locally as well, right? Which is really important for some industries too.

394
00:41:59,000 --> 00:42:06,600
And how did you handle the information architect chose it's RAC, a Genetic RAC or LMNVQ or

395
00:42:06,600 --> 00:42:12,840
how did you do it? Yeah, so it's kind of a Genetic RAC is part of the product. So we have

396
00:42:12,840 --> 00:42:20,440
Azure AI search which does the retrieval retrieval and then we show those results to an agent

397
00:42:20,440 --> 00:42:29,080
which is doing the RAC status and that's all done off like a tool call. So if the system recognizes

398
00:42:29,080 --> 00:42:32,600
that this is a piece of information, you're looking for a piece of information that it knows about

399
00:42:32,600 --> 00:42:36,920
it can go and do that. The nice thing about that is that we don't actually care where the information is.

400
00:42:36,920 --> 00:42:40,920
So we can do an API call, we can look for a SharePoint repository at that point, we could even look

401
00:42:40,920 --> 00:42:45,960
in your emails or something if that was a connector you want to switch on. But we can build custom

402
00:42:45,960 --> 00:42:53,240
connectors and have our agent look through the kind of different databases separately. And then we

403
00:42:53,240 --> 00:42:59,320
have separate agents to do the other tasks which are like taking a general kind of taking notes

404
00:42:59,320 --> 00:43:06,360
mode and report structure. So like please fill in what you've done today or what you've fixed.

405
00:43:06,360 --> 00:43:12,840
But the real benefit of them being a Genetic based is that you can flip from one to the other

406
00:43:12,840 --> 00:43:17,960
within the same conversation or the same session. So you can look up some information to put in that

407
00:43:17,960 --> 00:43:23,960
report really, really easily and it will just flick across the different agents. And I think also

408
00:43:23,960 --> 00:43:30,840
for your product important this computer vision and you are also the next product of your vision and also

409
00:43:30,840 --> 00:43:39,640
I think yeah, how fits it into your product. Yeah, it's that's something we're kind of like actively

410
00:43:39,640 --> 00:43:46,600
working on a lot. Because part, this, I would say this two sides to bit is one side which is incredibly

411
00:43:46,600 --> 00:43:53,960
easy, which is I have, I've got this model for you. Something like a serial number. Even with,

412
00:43:53,960 --> 00:43:58,520
so there are nightmare to write down, there are nightmare to read out because this I'd have to say

413
00:43:58,520 --> 00:44:06,520
dg20260 108. So it's long the Boris process. Some of them are like 20 Eric's long and people

414
00:44:06,520 --> 00:44:13,880
misread them or say numbers at the order all the time. So so easy for computer vision. Just take a photo

415
00:44:13,880 --> 00:44:19,720
of that, do an optical character recognition run, which is available in the Azure AI suite.

416
00:44:19,720 --> 00:44:28,280
And that's great. We can get, so we don't have to read that anymore. And that also applies to things

417
00:44:28,280 --> 00:44:33,160
like bar codes, QR codes, we can actually scan them in our app and get all the information out of them

418
00:44:33,160 --> 00:44:38,280
really, really quickly. That's not really machine vision to me. It's, it's, it is AI because we're doing

419
00:44:38,280 --> 00:44:44,680
things like OCR, but it's, we can do that on the device. We can do that using the, the power of the,

420
00:44:44,680 --> 00:44:50,040
the phone or tablet or whatever we're, we're doing the other side of it, which is much, much more

421
00:44:50,040 --> 00:45:00,520
complicated is, um, they, I was installing a window. And, um, my company that I'm installing that

422
00:45:00,520 --> 00:45:05,640
window for requires me to take a photo of the completed job. So I take a photo of the completed job.

423
00:45:05,640 --> 00:45:12,280
They want the AI to be able to say, you have installed that window correctly or you have not installed

424
00:45:12,280 --> 00:45:18,600
that window correctly. Now, there's some great amazing tools, some of them in Microsoft, some of them

425
00:45:18,600 --> 00:45:26,040
out of Microsoft to do that. It is perfectly possible. But we need thousands of examples. Maybe

426
00:45:26,040 --> 00:45:32,760
these days, we only need hundreds to do it kind of well. That gets difficult because we're then

427
00:45:32,760 --> 00:45:38,120
saying to our clients, please provide us with your last thousand window installs and we will make your

428
00:45:38,120 --> 00:45:45,480
own machine vision model. It's a little bit tricky. One thing we can experimenting with that we haven't

429
00:45:45,480 --> 00:45:51,320
implemented yet is actually, can you use a large language model, not a small language model in this case?

430
00:45:51,880 --> 00:45:59,160
Or it could be, I suppose, but a language model that can understand image inputs. Can you use that

431
00:45:59,160 --> 00:46:05,880
to get a description of what is in the image and maybe compare that description to what's expected?

432
00:46:05,880 --> 00:46:11,720
So for example, you could ask the LLM, can you see any large gaps between the freight, you could ask

433
00:46:11,720 --> 00:46:16,680
specific questions? And that to me is kind of like a middle ground approach. We're using really

434
00:46:16,680 --> 00:46:22,600
advanced image recognition and machine vision AI in a way that it wasn't quite intended,

435
00:46:22,600 --> 00:46:29,160
that that avoids us having to build these large custom models, which we are, I'm quite reluctant

436
00:46:29,160 --> 00:46:34,920
to do because the only is that another thing to kind of maintain, but it's somebody else's data

437
00:46:34,920 --> 00:46:43,160
and that doesn't really make it easy for us to do it. So the very, very basic solution we've

438
00:46:43,160 --> 00:46:50,040
gone for temporarily, whilst we figure that out is give us an example of one example of good,

439
00:46:50,040 --> 00:46:54,360
one example of bad and we'll show that to the user and we can use their brain to do it because

440
00:46:54,360 --> 00:46:59,560
we still have humans in the loop, so use them when necessary.

441
00:46:59,560 --> 00:47:12,440
I think the start for me is worse text. I think now we are living in the MoodyModal AI world.

442
00:47:13,400 --> 00:47:22,520
What did you think? How good are these models? And what you say there are people use it on

443
00:47:22,520 --> 00:47:31,400
their, and I don't know, smartphones. Yeah, how, I don't know, how good are these models,

444
00:47:31,400 --> 00:47:37,400
and I think they becoming bigger and bigger, how did you, can you run it on the smartphone?

445
00:47:38,280 --> 00:47:50,760
Yeah. So to me, sadly, they're a bit like cars, right? If you had a beta-pulled Ford Escort from 2004,

446
00:47:50,760 --> 00:47:59,720
maybe it's not the best car in the world versus a Formula One racing car that can do everything

447
00:47:59,720 --> 00:48:07,320
0 to 16, 0 to 16.1 seconds, whatever it is, there only is good as the driver. If you put a terrible

448
00:48:07,320 --> 00:48:11,560
driver in a Formula One car, they're going to crash. If you put a terrible driver in a Ford Escort

449
00:48:11,560 --> 00:48:15,960
from 2004, they're going to crash. If you put a racing car driver in either those, they're going to

450
00:48:15,960 --> 00:48:22,760
do better, but obviously the Formula One car is going to perform better because it was designed

451
00:48:22,760 --> 00:48:27,800
to do more. For me, it's exactly the same. I think the biggest problem that I see in other people's

452
00:48:27,800 --> 00:48:34,360
projects or like students that I'm mentoring or introducing to AI for the first time is they don't

453
00:48:34,360 --> 00:48:39,160
know how to prompt. I was really against prompt engineering when that came out as a phrase by the way

454
00:48:39,160 --> 00:48:43,640
I hated the phrase prompt engineering, but now I think it's a core skill. You have to be able to

455
00:48:43,640 --> 00:48:50,440
understand how the large language or small language models work in a vague capacity, what they're

456
00:48:50,440 --> 00:48:55,720
expecting and what they're going to output because their performance is so hugely dependent on

457
00:48:55,720 --> 00:49:00,600
the quality of the question that goes in. We have this phrase which is, shit goes in, shit comes out.

458
00:49:00,600 --> 00:49:08,680
It's so, so true because I'm constantly amazed by the accuracy and the performance that

459
00:49:08,680 --> 00:49:14,920
we're able to get out of these things if we show them and provide them the correct information.

460
00:49:14,920 --> 00:49:20,600
So actually the most valuable part of our product isn't necessarily the way that the AI generates

461
00:49:20,600 --> 00:49:27,560
and shows you an answer. It's about, that it gets the right answer in the first place in behind the scenes.

462
00:49:29,240 --> 00:49:38,760
I think that's really, really important. But having said that, so my last job I was working in research

463
00:49:38,760 --> 00:49:46,120
and development, a large consultancy firm, the performance of small language models today far

464
00:49:46,120 --> 00:49:50,920
outweighs the performance of large language models several years ago. So we are moving incredibly

465
00:49:50,920 --> 00:49:56,360
quickly, incredibly quickly. And there are language models you can run on on higher end

466
00:49:57,640 --> 00:50:02,680
devices and phones today. That are good enough for some jobs. So they're good enough for summarization,

467
00:50:02,680 --> 00:50:09,080
they're good enough for, you know, quick chat and messages. They will, I don't think they'll ever

468
00:50:09,080 --> 00:50:14,120
be good enough to do kind of a deep research, go off and find me 20 links and analyze the results,

469
00:50:14,120 --> 00:50:19,320
etc, but some of the, there's, I think it's, I kind of, yeah, it's the co-pilot research agent,

470
00:50:19,320 --> 00:50:23,160
right? So cool. Does loads. I don't think you'll ever have that running on your phone sadly.

471
00:50:24,840 --> 00:50:30,440
But I'm consistently impressed by the results when they used correctly.

472
00:50:30,440 --> 00:50:41,160
I think, yeah, AI, you say, shit, engine out and AI is the particular accelerator for this, I think.

473
00:50:41,160 --> 00:50:47,960
How do you handle governance and security?

474
00:50:49,000 --> 00:50:56,360
Yeah, so we're very reliant on Microsoft for governance and security. We have the platform

475
00:50:56,360 --> 00:51:01,400
completely hosted in Azure for exactly that reason. Microsoft put a hell of a lot more funding

476
00:51:01,400 --> 00:51:08,840
into security and a platform stability than I can at a very small company. So, you know, we,

477
00:51:08,840 --> 00:51:17,560
we rely heavily, heavily on them. We use a lot of their, you know, security services and

478
00:51:17,560 --> 00:51:23,320
monitoring services. It was very interesting to me, though, putting a company website up for the first

479
00:51:23,320 --> 00:51:28,600
time. This is the first company I've worked out where like I control the website and I can see who's

480
00:51:28,600 --> 00:51:33,400
trying to break into it and I can see the robots that are trying to pull information out of it. It's

481
00:51:33,400 --> 00:51:39,720
scary out there, man. It's, you know, my personal webbride gets like three hits a month and they're all

482
00:51:39,720 --> 00:51:47,480
from humans, but our company website is getting hit, you know, 100, 500,000 A by robots trying to extract

483
00:51:47,480 --> 00:51:53,880
end files from us and things. I'm just like, oh my god. Yeah. The Wild Western. Yeah. I think,

484
00:51:53,880 --> 00:52:02,760
yeah, I use Clouds, but they're really good. They're, yeah, it's really interesting. Also, I think

485
00:52:02,760 --> 00:52:12,920
actually 80% like the field worker part, 80% of all companies that are running on Clouds,

486
00:52:12,920 --> 00:52:23,800
it's also really interesting. Yeah, good. So, yeah, I could, yeah, we are really at the end, so I can ask you,

487
00:52:23,800 --> 00:52:32,200
I have so many questions we have to do in our podcast, I think. It's the, it's the cool, yeah, I have a

488
00:52:32,200 --> 00:52:43,240
rapid fire around. I ask questions and you give a, give a short, yeah, a short answer. So, okay,

489
00:52:43,240 --> 00:52:51,400
when you think, when you develop, what's the best drink coffee tea or energy drink? Energy drink.

490
00:52:53,640 --> 00:53:01,720
And, um, gaudy in the moron at the night, at the morning, at the night. I'm a afternoon

491
00:53:01,720 --> 00:53:05,640
coder. I don't work well in the morning and I don't work well in the evening. There's a sweet spot in

492
00:53:05,640 --> 00:53:15,080
the middle. What's your favorite arrow service beside Azure AI fondry? Sorry, what's my favorite?

493
00:53:15,880 --> 00:53:23,720
Azure service. Oh, it's, besides Azure AI fondry, it's got to be Azure AI search. I just think it's

494
00:53:23,720 --> 00:53:31,560
really, really good. And then, just, yeah, Azure hosted Kubernetes because there's nothing you can't do.

495
00:53:31,560 --> 00:53:40,680
What's your favorite AI model? My favorite AI model is 4.1 mini because I think it's just so

496
00:53:40,680 --> 00:53:48,600
flexible but is small enough to do stuff. The nano series are quite cool as well. My, I'll be cheeky

497
00:53:48,600 --> 00:53:53,240
and I'll give a second one, which is the, the GPT's image generation stuff. I think that's so cool.

498
00:53:53,240 --> 00:54:02,920
What, what did you think about AI help on, on, on, on programming?

499
00:54:02,920 --> 00:54:09,720
Really great tool if you know how to use it. If you know what you're doing, best thing in the world,

500
00:54:09,720 --> 00:54:15,000
if you have no idea what you're doing, a really great starter, a hugely problematic for people who

501
00:54:15,000 --> 00:54:21,800
think they know what they're doing. Is there one AI book YouTube channel or something you everybody

502
00:54:21,800 --> 00:54:34,680
should know about? That's a good question. I, my favorite AI video, it doesn't actually teach you

503
00:54:34,680 --> 00:54:40,920
very much but it's a generational neural network learning to play Mario. I just think it's such a good

504
00:54:40,920 --> 00:54:47,000
visualization of how computers really can learn over time by making mistakes. And I think that's

505
00:54:47,000 --> 00:54:51,720
really important visualization is like we only learn by making mistakes in the same as truth

506
00:54:51,720 --> 00:54:57,800
or computers. We just have to correct them. When you get an invite to the list bastards,

507
00:54:59,480 --> 00:55:07,640
if you're too far serious, what list do you boost? Our AI or in general.

508
00:55:07,640 --> 00:55:19,080
Okay, then also general. In general, this is crazy but I really want to build a laser system to

509
00:55:19,080 --> 00:55:24,920
hit the Apollo 11 laser reflectors on the moon because they left laser reflectors and everyone's like

510
00:55:24,920 --> 00:55:31,160
did really land. I'm like, you can do that. Grute yourself. Just get a really powerful laser. In AI,

511
00:55:31,160 --> 00:55:37,320
I would love to bust the myth that like AI is going to replace jobs. I just don't think that's

512
00:55:37,320 --> 00:55:42,120
where we are today. Maybe in the future it's just not good enough today. So let's think about how they

513
00:55:42,120 --> 00:55:47,960
augment people rather than replace it. There's something that Alia comes to you and say, hey,

514
00:55:47,960 --> 00:55:52,680
for us, you get all the money and resources you need. What feature will you develop?

515
00:55:52,680 --> 00:56:05,400
I, all the money and all the resources definitely robotic interaction with AI. I think we're doing so

516
00:56:05,400 --> 00:56:11,160
much amazing work with the kind of brains behind the systems tool calling best permits something

517
00:56:11,160 --> 00:56:18,440
more useful than Elon's human power to robots. Yeah, but it's really, really funny. I have started my

518
00:56:18,440 --> 00:56:29,880
career at the at the cook and we have done infrared scanning systems for for systems. Now,

519
00:56:29,880 --> 00:56:35,640
what's now happened? It's so amazing. It's really, really cool. Yeah. So

520
00:56:35,640 --> 00:56:44,840
what went when you say from this discussion, what is the thing the people should know

521
00:56:45,720 --> 00:56:51,320
or the only thing one information they can grab? What's the main information to be?

522
00:56:51,320 --> 00:57:00,360
Well, I always say to people at the end of such a thing, this is try it. Go on Azure, go on AWS.

523
00:57:00,360 --> 00:57:05,320
I don't care what it is. Go and try it. There's free accounts, there's free trials. All of these

524
00:57:05,320 --> 00:57:11,080
services have free tears. Go and try them. Go and use them because they're a lot of fun. And

525
00:57:11,080 --> 00:57:15,640
once you start trying and using and playing with these services, that's when the idea has come to you.

526
00:57:15,640 --> 00:57:22,120
That, oh, well, I see this problem in my real life and I see a solution here. So make the most of that.

527
00:57:22,120 --> 00:57:28,600
If you happen to be a student watching this, get your Azure student $120 credits. If you're not

528
00:57:28,600 --> 00:57:35,480
use the free tears, sign up, give it a play because you run way more by doing and playing than you do by

529
00:57:35,480 --> 00:57:44,920
any other method, I think. Okay. And finally, who will be the next guest on the podcast and what

530
00:57:44,920 --> 00:57:55,400
questions, shall I ask them? Interesting. I have two answers for you. One is our founder for

531
00:57:55,400 --> 00:57:59,720
Field Power AI, James Woodall, because he can come and talk to you all about the

532
00:57:59,720 --> 00:58:06,920
wearable compute world and how that came to be and why companies like Meta are not doing as well as

533
00:58:06,920 --> 00:58:13,560
companies focusing on the enterprise like these ones. And then I'll also tag in my fellow MVP,

534
00:58:13,560 --> 00:58:19,240
Josh McDonald, who can come in and talk about how, how Azure and Microsoft are actually securing

535
00:58:19,240 --> 00:58:24,760
stuff behind the scenes because he's a security MVP. I think I believe he's also an AI MVP now.

536
00:58:24,760 --> 00:58:29,480
And so he's a really knowledgeable guest around some of the questions that I kind of skinned over

537
00:58:29,480 --> 00:58:33,400
about security. I'm just like, believe it to Microsoft, just so we'll be able to explain that a little bit.

538
00:58:33,400 --> 00:58:38,600
Yeah, yeah, that's really good. Yeah, you have to make me the introduction then I am.

539
00:58:38,600 --> 00:58:45,960
And more brain in. Yeah, so thank you for joining me today on the podcast and sharing your insights,

540
00:58:45,960 --> 00:58:52,600
especially about AI, I found it, a genetic AI and computer vision and how, yeah, as an

541
00:58:52,600 --> 00:58:58,840
artificial intelligence can empower the front-line workers beyond it. Yeah, the traditional office

542
00:58:58,840 --> 00:59:10,920
workers. Yeah, this was the most amazing. And I think I don't know if I never expected I get someone

543
00:59:10,920 --> 00:59:18,680
in my podcast who built a master of. Let's carry out a piece of software for one of the camera subsystems.

544
00:59:18,680 --> 00:59:27,080
Yeah, so thank you. And yeah, this was amazing. Thank you so much.

545
00:59:27,080 --> 00:59:33,800
Thank you so much for the invite and yeah, hope to be back at some point.