Meeting Title: Uttam <> Mohamed Date: 2024-01-16 Meeting participants: Mohamed, Uttam Kumaran
WEBVTT
1 00:00:18.550 ⇒ 00:00:22.850 Uttam Kumaran: Hey? How’s it going? How are you?
2 00:00:22.900 ⇒ 00:00:26.090 Mohamed: Good? How are you? Good. Thank you.
3 00:00:26.870 ⇒ 00:00:27.729 Mohamed: How are you?
4 00:00:28.240 ⇒ 00:00:31.089 Uttam Kumaran: Yeah. Are you on the East coast.
5 00:00:31.140 ⇒ 00:00:33.350 Mohamed: II how about you?
6 00:00:33.860 ⇒ 00:00:54.969 Uttam Kumaran: Nice! I’m here in Austin, Texas. Oh, man, wait! I heard it was snowing in Arizona like couple of days ago. It’s like frosty, but I don’t know. It’s not that cool
7 00:00:55.350 ⇒ 00:01:01.800 Mohamed: but no it’s great to meet you from both Matthews.
8 00:01:02.240 ⇒ 00:01:26.899 Uttam Kumaran: Yeah, awesome. Yeah. And I’m and at the same as well. You know, I’ve been working, you know, speaking to them, you know, we we met just like exploring different ideas, and I’ve been helping them a little bit with their product. And yeah, just excited to kind of learn a bit about your background. And you know, I my background to give you context, I work in data engineering. I’m my background’s in computer engineering. And I work for the data engineer
9 00:01:26.900 ⇒ 00:01:50.420 for the past 6 years. And now II run my own firm now doing data development directly for clients. And then I’ve started to do some work in AI development as well. So kind of just on my own. II work in data engineering and led data teams before this. And yeah, just really was helping. You know both, Matthew, just think about the problem a little bit more that they’re trying to solve and think about how they’re building their product.
10 00:01:50.420 ⇒ 00:02:12.470 Uttam Kumaran: And yeah, just excited to meet new people and interested to see what you think after. You know, working with them for a little bit, but also about the product and hear about your background. So I mean, sure, yeah, I mean, that sounds great. Thank you for sharing all that. Yeah. So like my background was like physics, actually. But, like
11 00:02:12.570 ⇒ 00:02:19.710 Mohamed: decided to like, go into like data science for my master’s. And then lockdown and Covid happened so
12 00:02:19.820 ⇒ 00:02:37.560 Mohamed: biggest couldn’t be like can be choosers. So I somehow like found I found myself and like did engineering. So I was like doing that, for like the past, like 3, almost like 4 years. Now at Signa. So yeah, so it’s like, this is definitely like that element of like.
13 00:02:37.730 ⇒ 00:02:51.890 Mohamed: okay, I’m not particularly like scared of like data like data like, what it means is like work with like big data. But and like, there’s also like that element of working
14 00:02:52.020 ⇒ 00:02:58.520 Mohamed: cutting edge with also a bit too much sorry.
15 00:02:58.570 ⇒ 00:03:02.230 Uttam Kumaran: but you know, it’s like this is all kind of like very exciting
16 00:03:02.290 ⇒ 00:03:06.869 Mohamed: stuff that’s coming out like, especially the fi, like the past like few months.
17 00:03:07.000 ⇒ 00:03:10.150 Mohamed: Yeah, I think it is
18 00:03:10.480 ⇒ 00:03:18.319 Mohamed: quite exciting to like, just like, be at this like point in time and like, be able to like work on all this as this is actually happening because
19 00:03:18.560 ⇒ 00:03:21.060 Mohamed: looks like it is actually finally happening.
20 00:03:23.090 ⇒ 00:03:43.070 Uttam Kumaran: Yeah, II feel the exact same way. You know, I think a lot of the stuff in AI and things like that is happening. And the tools are really cheap and available. So kind of the scenario somewhere, I said, the main thing. Now, it’s really good to be like the intersection of data as well. And again, I’m a developer. So for me, it’s like very easy for me to like conceptualize a lot of the ideas. And then
21 00:03:43.160 ⇒ 00:03:51.529 Uttam Kumaran: think about okay, like, how fast can we think we could build this thing or like, what are the key components? Like, what do you think about their idea. And like, have you worked with?
22 00:03:51.720 ⇒ 00:04:00.979 Uttam Kumaran: Have you? Have you done any like AI prototyping stuff in the past, or and also like, what do you think about their overall idea. Yeah, I started talking to them when when they were thinking about it.
23 00:04:01.030 ⇒ 00:04:14.029 Uttam Kumaran: Kind of like, you know, helping, trying to do like biometrics and things like that. And they kind of pivot to kind of really focus on summarizing for sales. And, you know, kind of starting there, but interested to hear your thoughts after taking a look and trying stuff out.
24 00:04:14.410 ⇒ 00:04:19.750 Mohamed: I mean, sure. Yeah. I mean, having like looked at everything. I think
25 00:04:20.860 ⇒ 00:04:40.639 Mohamed: it is like kind of like a very like exciting time. But it also like means that everyone else is just like sort of like rushing in with fairly like similar ideas or like, there’s like a lot of like overlap in terms of functionality between this thing with that thing definitely like an over abundance of like choice at this point.
26 00:04:40.790 ⇒ 00:04:52.729 Mohamed: So I think execution is kind of like key to this. I this like product more than like more than anything else more than the actual like idea, like, how good the idea is.
27 00:04:52.820 ⇒ 00:05:01.820 Mohamed: cause, like, essentially, yeah, like, this has been done in some form of or another in the past. So
28 00:05:01.880 ⇒ 00:05:06.670 yeah, so doing it. Well, I think that’s like, sort of like, where it’s actually kind of key.
29 00:05:06.970 ⇒ 00:05:17.329 Mohamed: And what does that actually entail. Well, for me, it’s like, obviously, it’s like that. Ha! That means like having, like a robust and scale ones like
30 00:05:17.880 ⇒ 00:05:20.320 Mohamed: streamlined like back-end system
31 00:05:20.360 ⇒ 00:05:23.700 where we have, like the connectors in place.
32 00:05:24.050 ⇒ 00:05:27.910 Mohamed: sifting the data through whatever data source that we want.
33 00:05:28.550 ⇒ 00:05:45.139 And then sec, so dumping them into like postgres, like the database, for example, where each scheme is a different data source, for example, and within that and within that like an every database is a different client. For example, what have you?
34 00:05:45.330 ⇒ 00:05:50.499 Mohamed: And then being able to like, actually like integrate
35 00:05:50.520 ⇒ 00:06:03.699 Mohamed: python, python functionality, or less like computational functions alongside, leveraging like open ais like chat. Gpt, I think that’s like sort of
36 00:06:03.850 ⇒ 00:06:06.410 Mohamed: very exciting.
37 00:06:06.620 ⇒ 00:06:07.710 so
38 00:06:08.060 ⇒ 00:06:14.319 Mohamed: being. And then I think the the end goal, or like at least like the the main
39 00:06:14.890 ⇒ 00:06:20.030 Mohamed: the main output would be like presented to the users through the use of a slack bot?
40 00:06:20.230 ⇒ 00:06:37.290 Mohamed: so yeah, like, I think that was like, sort of like a a very quick overview, I can like tell, like I can elaborate a bit more, or if you wanna like, chime in at any point, please let me know. So one thing I’m doing in my company is like, I just have everything hooked up to Zapier, and like I’m
41 00:06:37.700 ⇒ 00:06:51.549 Uttam Kumaran: having zoom record everything, and then I have it go through zap here, and it stop me so like that’s one thing like I think you could probably honestly get really far like a prototype just through like messing around on Zapier. But I think the tougher part is like
42 00:06:51.700 ⇒ 00:07:00.989 Uttam Kumaran: having observability on that platform. And then also, that’s like this is that’s like, just for internal like, I’m not building a product or anything so definitely like, I don’t know if you’re thinking about running
43 00:07:01.020 ⇒ 00:07:22.349 Uttam Kumaran: like I had never used. I, Matt told me, like it’s all fireflies and things like that, and I don’t know about the Api. But I assume you could just like hit their Api get recordings and then process stuff like, do you get transcripts or like, what do you get so like, I actually like pulled up? There’s like a couple of like more click transcripts. So I actually like pulled
44 00:07:22.650 ⇒ 00:07:23.840 Mohamed: data
45 00:07:24.040 ⇒ 00:07:34.870 Mohamed: from like the Api, and I pull data from like a postgres like, I set up a Postgres server. I connected to it through using that, and was able to like fetch records from that as well.
46 00:07:35.120 ⇒ 00:07:41.829 Mohamed: So the graph, like the the graph Ql. Api, I mean.
47 00:07:42.150 ⇒ 00:07:51.689 Mohamed: that was like, sort of like, not a point of contention. But it was definitely like a conversation about like, Okay, what are we using for the connectors. What are we? How are we getting it there?
48 00:07:51.770 ⇒ 00:08:07.160 Mohamed: To me it definitely like makes sense to sort of like, take take off like the heavy lifting and like areas where we don’t need to like excel. Essentially or like, you know, standard is fine.
49 00:08:07.580 ⇒ 00:08:16.220 Mohamed: And that’s like what I meant about like what I think would dis like the distinguish this particular project more than anything else. Is not
50 00:08:16.410 ⇒ 00:08:17.980 Mohamed: using
51 00:08:18.130 ⇒ 00:08:26.170 Mohamed: zapier, or arch, or this thing with that thing. It’s it’s actually just like being able to like.
52 00:08:26.850 ⇒ 00:08:31.030 Mohamed: construct the workflow that is as useful to a user as possible.
53 00:08:31.040 ⇒ 00:08:33.370 And at this point.
54 00:08:33.480 ⇒ 00:08:46.239 Mohamed: I think what we sort of like settled on is essentially having like a bot that would replicate entire like users like workflow and like, provide them like with notches like throughout.
55 00:08:46.610 ⇒ 00:09:04.889 Mohamed: So if you, if a meeting has just like finished, you send like the transcript over like it gives you like a few action points testing that thing. If it has like access to calendar, they could suggest a time. I think, like that, actually like presents like
56 00:09:05.380 ⇒ 00:09:16.420 Mohamed: to do opportunities to like grow the product beyond just like sales managers or like sales employees, just like, okay. If everyone has, like their own, like super smart psychic at work. That’ll be
57 00:09:16.560 ⇒ 00:09:29.750 Mohamed: yeah. So I think, like, that’s sort of like, where we sort of are different from everyone else or like, we’re like, try, I think that’s like where we’re like going to like, try and like, focus most.
58 00:09:30.140 ⇒ 00:09:34.990 Mohamed: So yeah, back to like sort of connectors and Zapier’s. I think
59 00:09:35.170 ⇒ 00:09:47.960 Mohamed: Zapier is also like launched like a new like table function. So so like essentially data warehousing. I don’t know how. Yeah, a robust scalable that is, but it’s it’s there, at least for now.
60 00:09:48.130 ⇒ 00:10:04.960 Mohamed: And like, it seems to me that they’re like integrating the system are like fairly heavily with open ais like Api, they’re trying to like be the main. Yeah. So like, there’s obviously like plans for like companies and enterprises. So
61 00:10:05.240 ⇒ 00:10:12.849 Mohamed: besides, like cause, like, for example, one of the the requirements of the prototype is to like, have
62 00:10:13.280 ⇒ 00:10:20.000 Mohamed: a meeting transcript be generated whenever a new meeting is like finished. Besides, like, how like utilizing Zapier’s triggers
63 00:10:20.170 ⇒ 00:10:26.150 Mohamed: it might be like a little bit more trickier to like. do you know?
64 00:10:26.230 ⇒ 00:10:32.959 Mohamed: So I’m actually curious. If it weren’t like, if it if it if you went to the queue, zip it in that situation, what would you do?
65 00:10:33.740 ⇒ 00:10:43.930 Uttam Kumaran: Yeah. So open. Api open, has, like their whisper, Api, where you pass in pretty much like audio, so like I haven’t. I don’t know what you get back from fireflies, but
66 00:10:43.960 ⇒ 00:11:08.520 Uttam Kumaran: pretty much like I don’t know. She probably seems best like with that Api and I don’t know where you’re wherever you’re running this. But somewhere on you, machine or something you hit the Api get the latest. I don’t know if you can get again. If you can get a link to a file, or if you can get a transcript, then you could pass. Then you could pass that. If you’re just getting a filed audio, then you could pass that directly to
67 00:11:08.630 ⇒ 00:11:37.009 Mohamed: like generates both. So there’s like it. Stores like the audio file, and also like transcribes it. So like, even in like the transcription file, or like, when I pull the data at the end of the at the dictionary also includes, like a link for them to to do for yeah. And like, right now, there’s like, I have a function that sort of like fetches it converts it to Json and like stores it in a particular directory.
68 00:11:37.090 ⇒ 00:11:43.329 Uttam Kumaran: The idea. He is like sort of like that access easily by any sort of agent or Llm model.
69 00:11:43.540 ⇒ 00:11:44.900 Mohamed: Yeah.
70 00:11:45.290 ⇒ 00:11:46.960 Mohamed: Duh.
71 00:11:47.560 ⇒ 00:11:58.429 Uttam Kumaran: I mean? That’s probably like the the key thing like again. I don’t. I don’t know how important is to figure out like where it gets hosted or what lm, to use like, I think the biggest thing is like, can you take any
72 00:11:58.460 ⇒ 00:12:19.970 Uttam Kumaran: text transcript for any audio? Right? And that way you can integrate with like fireflies, or zoom, or whatever. Can you then pass that to lm through like either laying chain or something with like templated prompts, right like whatever the key thing have those set up?
73 00:12:20.510 ⇒ 00:12:35.079 Mohamed: I tried to like, create like a a, a function that would couldn’t like translate. But like movies like for like for our looking movies. It’s just like you can’t find like a lot of out ofic movies with subtitles. It’s like, I wanna like, show them to people
74 00:12:35.160 ⇒ 00:12:38.690 Mohamed: like. For some reason it all came out in Russian.
75 00:12:38.900 ⇒ 00:12:56.100 Mohamed: but I think, obviously with like English audio, I think it can go for like a far way. So it would be like, actually interesting to see the difference and results between using fireflies
76 00:12:56.210 ⇒ 00:13:00.409 Mohamed: or automatic like transcribe text versus utilizing whisper.
77 00:13:00.730 ⇒ 00:13:14.910 Uttam Kumaran: I think Whistler is probably gonna be better because Whistler is really good, like even the Zoom transcripts. Sometimes there are messed up like, for example, like sometimes people’s names aren’t correct. Like, there’s there’s issues.
78 00:13:15.180 ⇒ 00:13:28.899 Uttam Kumaran: but also you can also get rid of like stutters like you want to get rid of stutters. Get rid of filler words and things like that. You could also add metadata. Right? It’s like, if if the person tells you this, call the goal is this.
79 00:13:29.130 ⇒ 00:13:30.899 Uttam Kumaran: then it kind of may change.
80 00:13:30.960 ⇒ 00:13:35.360 Uttam Kumaran: So I’ve had calls with this like small talk and stuff. I’m like, get rid of that. Nobody cares about that.
81 00:13:35.370 ⇒ 00:13:42.689 Uttam Kumaran: We’re talking about the weather, or something like that. I don’t. I doesn’t need to. I don’t want that to skew any results. This call. The purpose was to do this.
82 00:13:42.760 ⇒ 00:13:46.560 Uttam Kumaran: And then, like, I want a summary of like, what? What are the action items?
83 00:13:46.580 ⇒ 00:14:07.670 Uttam Kumaran: And then I want to be able to share that real quickly, right? And then I want that in the email, I want them slack. So you’d be able to say like, so there’s almost some phone pieces like grabbing all that stuff information being able to pick what Lolm or what transcription model is the best producing the summaries and then routing the summary somewhere. But on either end. Those are all like kind of like back end.
84 00:14:08.270 ⇒ 00:14:13.759 Uttam Kumaran: Those are like backend, kinda like back end problems, not really like tough data problems.
85 00:14:13.820 ⇒ 00:14:37.490 Uttam Kumaran: I think the data problem is like, what you’re saying is like storing it. But then, also doing like this sort of like prompting and like figuring that stuff out that if you could isolate and get right, then I think you could scale that. However, like I don’t think it doesn’t. That is the part where I feel like is the toughest thing is like, can you reliably get summary and things like that? You’re not getting like nonsense or like garbage stuff.
86 00:14:37.520 ⇒ 00:14:45.199 Mohamed: I mean, so far like, obviously, you can go for far away with like fine tuning like the prompts and all that.
87 00:14:45.250 ⇒ 00:15:00.149 Mohamed: Right now, I’m just like working on testing out instead of like using phones. It’s like a music like utilizing agents. So have like a bunch of agents, sort of implement checks and balances for one another.
88 00:15:00.270 ⇒ 00:15:05.979 Mohamed: The idea here is like, you’re essentially digitizing every
89 00:15:06.350 ⇒ 00:15:16.449 Mohamed: person at the company, and it’s like there’s a CEO, whatever and then, you know, it’s like you do that you have them like collaborate on like a bunch of shit, and
90 00:15:16.490 ⇒ 00:15:34.500 Mohamed: you end up with like with the best answer, essentially but what I wanted to ask you again is like a sort of do you see anything missing in that like workflow like that sort of pipeline so far. Do you see any any areas where we might be missing?
91 00:15:35.820 ⇒ 00:15:47.799 Uttam Kumaran: Yeah, I mean, I think you should number. I think the the thing that I’ve read a lot about is just like the the effectiveness of metadata. So you should certainly try and take the the name of the event.
92 00:15:47.980 ⇒ 00:15:54.910 Uttam Kumaran: and any ancillary information that the user can provide about who they’re meeting. The purpose of the meeting.
93 00:15:54.960 ⇒ 00:16:13.929 Uttam Kumaran: I think it’s actually, although it seems small is not really sometimes clear from the contents of the meeting. I mean, the second thing is also the meeting clean up. So being able to remove a lot of filler words removing like stuff. That’s unnecessary. I think that’s actually super super key. And like the sort of like pre processing the files.
94 00:16:14.060 ⇒ 00:16:18.889 Uttam Kumaran: I just I just think in my experience, if the Lms will take a bunch of like.
95 00:16:18.940 ⇒ 00:16:34.639 Uttam Kumaran: for example, from Zoom, I’m getting like action items about like the small talk section and like, I don’t. If you should be able to clearly identify that that’s not related to the call and get rid of that. And and so I think there’s probably a lot of work to be done just on the pre processing side.
96 00:16:34.750 ⇒ 00:16:41.399 Uttam Kumaran: So I would say that. And then the second thing is evaluation. So the one thing II talked a lot of people about is like.
97 00:16:41.460 ⇒ 00:16:46.820 Uttam Kumaran: Yeah, you may feel that these thing like we’re getting closer. But I think the best thing is to do is like, look into
98 00:16:46.880 ⇒ 00:17:07.009 Uttam Kumaran: link chain, and some other stuff has a great like eval framework to like how to do evaluation and trying to put into place a measurement of like is this, is this actually like system getting better and like, are we getting the output. So we is. Then you can scale that up and run tests and kind of that. But I think implementing like that sort of evaluation early is probably
99 00:17:07.210 ⇒ 00:17:25.580 Mohamed: for sure, like, obviously, when it comes when we’re like down to like building stuff. You need testing is like, obviously like a must like, even like when I was like collecting like the Apis, or whatever it’s like writing like an auxiliary function to just like check the Api connection status. I think that’s like, sort of like, very important.
100 00:17:25.650 ⇒ 00:17:39.170 Mohamed: But yeah, on your like. First point as well. It is like like again, like an exciting time, because, like with the G chat, Gpd, for, like the increase like token size.
101 00:17:39.180 ⇒ 00:17:45.750 Mohamed: it’s it can essentially like fit, like a 300 page book into context and like before it was.
102 00:17:46.060 ⇒ 00:17:47.849 Mohamed: what was it? 16 or some shit?
103 00:17:47.860 ⇒ 00:18:03.210 Mohamed: So yeah, that that would definitely like mean, like, a substantial improve in performance at a perspective that’s like kind of like, why, you know, like testing out the different like, cha, like versus 4,
104 00:18:03.370 ⇒ 00:18:05.420 Mohamed: it yeah fits.
105 00:18:05.880 ⇒ 00:18:11.690 Mohamed: You’re very like, almost pretty much like it. That’s sort of like, okay, how can we?
106 00:18:11.990 ⇒ 00:18:20.019 Mohamed: If everyone has the same output? How can we sort of like distinguish ourselves, or like our product in a way where where we’re just like not really like
107 00:18:20.180 ⇒ 00:18:28.689 Mohamed: opening a window for Gpg, just like you’re trying to like come up with like functionality that actually come like it enables it a little bit better.
108 00:18:29.100 ⇒ 00:18:52.170 Uttam Kumaran: Yeah, I think that’s probably gonna come up with like being able to constantly reference meetings over like again, as you save all those meetings and the outputs you can now understand, like, Hey, you’ve like historically acted in this certain way, or you know, I think there’s probably some benefits to get being like the people that have analyzed all the certain persons meeting, and you almost have a persona bill
109 00:18:52.170 ⇒ 00:19:10.229 Uttam Kumaran: for that, you know, and you could give people more feedback over time, right? But it’s that’s probably like one of the bigger things. And then, like, I don’t know. Yeah, yeah. So like again, as as like, for example, I I’ve been recording almost all my meetings for the past like 4 months, and my that’s my plan is to like, have
110 00:19:10.230 ⇒ 00:19:19.899 Uttam Kumaran: it analyze not only the way I speak, but my audio, and be able to almost replicate or give me like what I would decide to do in certain situations.
111 00:19:19.900 ⇒ 00:19:43.500 Uttam Kumaran: And I realized that, okay, the goal is just to record every single conversation that way. I will build that over time. So it has context of the way I make decisions. And so I think things like, that is stuff where, yeah, if you just open like something on the Gpt store. It’s gonna be okay. But you want something that has all your contacts. Also your company’s contacts or team members context. So there may be again some some stuff to do on like the Pre processing collection side
112 00:19:43.670 ⇒ 00:19:53.439 Uttam Kumaran: that like, if you were just giving a transcript, maybe you can get like 50% fidelity. But if you knew the company was this, if you knew, the goal of the meeting was this, if you knew the person’s background was X,
113 00:19:53.450 ⇒ 00:20:03.909 Uttam Kumaran: like, maybe able to get way farther. That’s, I think, the biggest differences between like something off the shelf and something that’s like really integrated, and you may be able to get a lot of that from their slack.
114 00:20:03.930 ⇒ 00:20:17.590 Mohamed: But I mean, yeah, that’s like, it’s sort of like describes what I would want to like have like happen like underneath the. So that’s like sort of like the main. I main premise behind, like having those like agents
115 00:20:17.600 ⇒ 00:20:45.650 Mohamed: I like I would want like ha! I would want to like have like a dedicated agent to like go through like Csv files. I want like to have someone kind of like like summarizing meeting minutes I want them to like all like be reviewed by one person’s like having all that happen. I think that’s like very important. And I think, yeah, so like regarding like the preprocessing side. You know, like.
116 00:20:45.690 ⇒ 00:20:49.320 Mohamed: there’s definitely like, weird. Talked a lot about
117 00:20:49.450 ⇒ 00:21:11.240 Mohamed: it it’s all like well and good to sort of like have a super shiny like chat like application. But if we can actually sort of like, take off, take take away some of the heavy lifting with like actual like computation that would like drastically like either, like
118 00:21:11.650 ⇒ 00:21:15.270 Mohamed: decrease costs, because, like, you know, utilizing the Api.
119 00:21:15.310 ⇒ 00:21:17.920 I might might just like actually be fasted.
120 00:21:18.070 ⇒ 00:21:29.100 Mohamed: So, for example, like one thing that we discussed is after. Like, we pull, like all of the data and put them into like the postgres database we can create like some sort of like master user table.
121 00:21:29.240 ⇒ 00:21:40.079 Mohamed: that kind of like references. All of those different data sources. Can include also information that we like feed it from the actual user and interaction.
122 00:21:40.190 ⇒ 00:22:09.710 Mohamed: I think having like, yeah, having like character profiles are like logs, or like personas of PE, like digital personas of people would go a long way, especially if you’re in a managerial role, if you not only have, like a meeting, or like transcription, minute minutes, or whatever or like action items you also like, know what? John is like, what is like, what this person is like, you know? So I think.
123 00:22:10.130 ⇒ 00:22:20.489 Uttam Kumaran: yeah, there’s a lot to be done. It’s exciting. But yeah, integrations. Right? It’s like on the sales side. There’s so many Crm system, so many like
124 00:22:20.700 ⇒ 00:22:32.419 Uttam Kumaran: sales logging systems at the top part is like, have integrations with all of them. So that I don’t know. That’s like a yeah. How do you see the connective like aspect of it?
125 00:22:33.690 ⇒ 00:22:51.509 Mohamed: Do you think, in terms of setting that data instead of sending it out or no, in a second terms of like better to like, build it out ourselves, or to just like have like, try and get the most out of Zapi at and or another like data warehousing company, or like.
126 00:22:51.790 ⇒ 00:22:59.329 Mohamed: yeah, it’s tough. I mean, I would probably make sure you at least have integrations with the top 3 like Crms.
127 00:22:59.390 ⇒ 00:23:10.149 Uttam Kumaran: like, there’s a ton of them. But again, most people use Hubspot salesforce, and probably like one other. So like really making sure that you have those integrations.
128 00:23:10.220 ⇒ 00:23:31.529 Uttam Kumaran: and that like that. Those exist, because, again, like, it’s easy for me individually to go build just a quick black bot for my company. But I can’t build integration to. It’s like, that’s where it gets complicated. That’s what people are gonna immediately ask for is like, how do you action on that data? So you’re gonna say, like, Oh, I met with this customer. That customer then gets updated in Hubspot
129 00:23:31.640 ⇒ 00:23:49.049 Uttam Kumaran: with, like the action items and things like that. So those are the deeper things that now actually begin to save a lot of time. That’s what I think is like, I don’t know technologies that offer that easily. But that’s like, that’s the stuff that’s
130 00:23:49.190 ⇒ 00:23:50.980 Uttam Kumaran: Yeah, they’re safer. But
131 00:23:51.030 ⇒ 00:23:56.630 Uttam Kumaran: Zapier is. If you’re if you’re sitting on top of Zapier, it’s a huge cost.
132 00:23:56.760 ⇒ 00:24:01.740 Uttam Kumaran: like it’s if you’re running if you’re running like hundreds of thousands of requests.
133 00:24:02.330 ⇒ 00:24:06.750 Uttam Kumaran: Thank you. Really 1,000 requests per day to to do small updates.
134 00:24:06.920 ⇒ 00:24:17.419 Uttam Kumaran: you end up paying them quite a bit of money, but some of some of the some of the applications you can write like you can. You can write Api’s The Hubspot and things like that. It just again, you’re writing software. So
135 00:24:17.470 ⇒ 00:24:25.699 Uttam Kumaran: that’s another thing to maintain. And things like that. So I don’t know. I would say, like, see what the first customer you guys are going after like.
136 00:24:25.720 ⇒ 00:24:41.550 Uttam Kumaran: what do they use for the Crm. Where do they spend their time like logging deals and saying like, Hey, if we if we not only gave you summaries, but we also updated the hubspot. Does that save you some amount of time? That’s the thing is gonna be, there’s a lot. There’s a ton of these different Crm applications that
137 00:24:41.550 ⇒ 00:24:58.750 Mohamed: exactly. I mean, that’s sort of like it. And like having to like, delve into like each and every one’s like documentation, and like this thing or that thing, and figuring out for yourself. It like that, is like an added like, I spent easily 70 80% of like the time working. So far, it’s like
138 00:24:58.800 ⇒ 00:25:06.839 Uttam Kumaran: going through like the graph. Ql, like a like documentation is like, okay, like, it’s just like, sort of mundane work. But like, it needs to be done somehow.
139 00:25:06.940 ⇒ 00:25:30.570 Mohamed: but yeah, no, that’s that’s interesting. Yeah, that kind of like already built in like tools, for, like Api retrieval.
140 00:25:30.710 ⇒ 00:25:47.840 Mohamed: in terms of like how good it is or like how scalable it is. I’m not too sure, honestly. But yeah, like having yeah, having the connectors like them. That’s it. That would be a great, great idea.
141 00:25:48.030 ⇒ 00:26:15.980 Uttam Kumaran: But then, again, it’s like you. You may want to integrate with discord slack teams. There’s there’s a ton of difference. So it’s again, it’s like a market share thing. So just making sure that, like you have again, if you have a if you have your ability to actually take in any transcript and produce that message, then you then you can have someone on the back end team focused on sending that to teams or whatever, and then focusing on sending it. All these things like the hardest part here is not that part has already been done. Many companies are doing that hardest part is this sort of like
142 00:26:15.980 ⇒ 00:26:31.099 Uttam Kumaran: AI summarization and producing that stuff is like, what’s I like? That’s the the bulk of the issue. I think if you guys improve like that works, then you can get somebody to come in and do other integrations
143 00:26:31.140 ⇒ 00:26:39.339 Mohamed: so like in terms of like development, to like just taking this to production like, how do you see this like sort of like going from like a high level.
144 00:26:39.440 ⇒ 00:26:44.209 Mohamed: just want to like, get your like inside of like, input, because.
145 00:26:44.330 ⇒ 00:26:47.889 Uttam Kumaran: yeah, I mean, II would.
146 00:26:48.940 ⇒ 00:27:14.279 Uttam Kumaran: Yeah, I mean, I don’t know. I would. I would pick like where all your process are running. I mean, you could. Just. You don’t need anything. You really just need, like a post based database for processing files and platforms. You can just put all that in. I would likely pull all the recordings, put all that Ms. 3, and then start writing functions to pull those that way. You don’t have to hit the fireflies. Api again. You have all that, and then again, you just
147 00:27:14.280 ⇒ 00:27:31.559 Uttam Kumaran: build like your basic database of like here, like, here’s a transcript. Here’s the summary, and then you built. Then you put your building, and you need all the logic sitting somewhere. So I would have a lambda or something where you could call that back end logic, which is like all the Lm logic. And then
148 00:27:31.670 ⇒ 00:27:51.109 Uttam Kumaran: again, you. This application at the moment doesn’t sound like it has much of a front end, so everything kind of has to happen on the trigger or kind of on a function call no like I. So like, I initialize it and then so I can actually show you real quick if you’d like.
149 00:27:51.320 ⇒ 00:27:53.420 Uttam Kumaran: yeah,
150 00:27:53.620 ⇒ 00:27:56.250 Mohamed: And since I created
151 00:28:00.320 ⇒ 00:28:04.969 Mohamed: sort of like in the thick of it at the moment, so apologies for the mess.
152 00:28:05.800 ⇒ 00:28:11.780 Mohamed: So, yeah, like, this one is actually coming through from postgres.
153 00:28:12.260 ⇒ 00:28:15.629 Uttam Kumaran: right? And this one is actually coming through from like
154 00:28:16.490 ⇒ 00:28:23.150 Mohamed: from graph, like the Api, and like this is what the actual like trends transcript. Looks like.
155 00:28:25.750 ⇒ 00:28:26.630 Uttam Kumaran: yes.
156 00:28:26.820 ⇒ 00:28:30.770 Uttam Kumaran: okay, great. Oh, nice. Well, that’s great.
157 00:28:30.960 ⇒ 00:28:44.669 Mohamed: So it’s like, not too bad. It’s like you have, like the start, time and time. Who’s speaking like in the metadata is like actually like being fashion. There’s like a fair amount of, like other metadata that you can extract. You get all the text, pass it
158 00:28:45.040 ⇒ 00:28:50.230 Uttam Kumaran: you like, unravel it, get all the text, pass it down, and then. yeah.
159 00:28:50.290 ⇒ 00:28:58.499 Uttam Kumaran: quickly. Just prompted to say, like, Tell me what
160 00:28:58.660 ⇒ 00:29:01.140 Mohamed: I can either. Just like
161 00:29:01.460 ⇒ 00:29:12.709 Mohamed: I feed it like the actual like text that I’m like feeding into the slack bot as well as like the database. The data, the database bullshit. Sorry
162 00:29:12.880 ⇒ 00:29:14.759 Mohamed: while system synchronization.
163 00:29:15.680 ⇒ 00:29:16.840 Mohamed: I don’t change anything
164 00:29:17.130 ⇒ 00:29:23.709 Mohamed: or and it’s like, feed it like the message that I’m feeding into it. And like everything that’s coming from postgres and like
165 00:29:23.740 ⇒ 00:29:24.830 Mohamed: everything else.
166 00:29:24.990 ⇒ 00:29:27.450 Uttam Kumaran: So it is like exciting stuff. Yeah.
167 00:29:28.630 ⇒ 00:29:40.710 Uttam Kumaran: I mean, it looks like it’s at least it’s like working pretty well. I think the biggest thing is again like making sure that you could do like. I think it’s good at like these sort of broad tasks. The problem is, it misses stuff.
168 00:29:40.870 ⇒ 00:29:59.549 Uttam Kumaran: So the tough part with like stakeholders is like, if it misses an action item, then like, what do you do? Right? Because sometimes I’ll look at the summary, and I’m like. I just don’t know what they got everything. So it’s like, how do you make sure that it has everything, or that the people can get feedback or things like that?
169 00:29:59.830 ⇒ 00:30:14.899 Mohamed: That’s key? Because again, if if you have like 10 action items and gets 8, I don’t know whether that’s like that’s I don’t know whether that’s success for that, I suppose, would be like process, both like text and raw audio
170 00:30:15.240 ⇒ 00:30:29.040 Uttam Kumaran: but less about missing less about missing it, more like that. The language model just like doesn’t think it’s to do different models. Look at it.
171 00:30:29.750 ⇒ 00:30:36.580 Mohamed: See? Yeah, it’s like, sort of like another pair of eyes to see if like this one missed anything.
172 00:30:36.940 ⇒ 00:30:58.189 Uttam Kumaran: So that’s the stuff where I think, if you have like, if you have a if you have a couple of examples already, and you have the evaluations like, for example, if you’re like, here are the 10 action items. We should always be hitting these like. That’s a good way to just evaluate whether those are working. And then again, what you can also do is if you’re running stuff on Github, you can run like github workflows to run these tests.
173 00:30:58.200 ⇒ 00:31:06.259 Uttam Kumaran: you know, just to say, like, I, I’m running this code. The output should be this like or things like that. And then again, you can even have agents that
174 00:31:06.380 ⇒ 00:31:18.119 Uttam Kumaran: that can examine like, Hey, this is output. Pretty much match, for example, output. That’s some of the evaluation frameworks that again, for someone like probably like Matt, or even a client for you to be like. You’ll be expecting
175 00:31:18.380 ⇒ 00:31:26.300 Uttam Kumaran: that we expect the answers to be even like once you make a change, too, because you may make a change to the prompt, or something that you don’t really know.
176 00:31:26.550 ⇒ 00:31:51.550 Uttam Kumaran: Like, what’s the downstream effect entirely. So that’s the sort of stuff is like. That’s where I try to rely on the kind of testing. So I don’t cause I don’t. Wanna. I don’t wanna say that I just wanted to be really super clear that like same output, we get the same out, but we get the same like pretty much output, like, no matter what we do. That’s not that I think that’s a spot on like that. That’s sort of like what I’ve been trying to like pitch to. Matthew.
177 00:31:51.550 ⇒ 00:32:13.319 Mohamed: it’s like trying to like, create like modules, or like almost dashboards, so that we can sort of like in calculate whatever we need to calculate, and whatever actually needs a large language model. We utilize it then. But anything else like, I don’t wanna like crunch numbers with like a large number large language model, if I if I can help it.
178 00:32:13.370 ⇒ 00:32:16.730 Mohamed: it’s like, sure. Yeah.
179 00:32:16.750 ⇒ 00:32:19.500 Uttam Kumaran: yeah, let me even try to send you this like.
180 00:32:19.550 ⇒ 00:32:23.880 Uttam Kumaran: evaluation thing on Github that I like.
181 00:32:24.440 ⇒ 00:32:27.340 Uttam Kumaran: But I reference a bunch.
182 00:32:30.780 ⇒ 00:32:32.290 Uttam Kumaran: that’s okay.
183 00:32:35.710 ⇒ 00:32:37.050 Uttam Kumaran: Yay.
184 00:32:37.820 ⇒ 00:32:38.500 Oh.
185 00:32:43.250 ⇒ 00:32:46.409 Mohamed: like even as Zoom is offering that
186 00:32:46.870 ⇒ 00:33:04.999 Mohamed: the choice to like send an automatic somebody to like all participants. So it’s like, yeah, yeah, I don’t know how far they’re gonna be able to get.
187 00:33:05.200 ⇒ 00:33:10.490 Uttam Kumaran: There’s one other one and see that
188 00:33:23.140 ⇒ 00:33:24.710 Mohamed: what’s called Travis.
189 00:33:27.730 ⇒ 00:33:29.549 Uttam Kumaran: Yes.
190 00:33:30.680 ⇒ 00:33:33.870 Uttam Kumaran: Oh, it’s called. I think it’s probably
191 00:33:38.520 ⇒ 00:33:42.679 Uttam Kumaran: yeah. So these 2 are probably interesting to take a look at.
192 00:33:45.680 ⇒ 00:33:49.880 Mohamed: I looked at one that like sort of like had like an economic
193 00:33:50.140 ⇒ 00:33:58.060 Mohamed: framework for agents use like, I kind of like that one I was like.
194 00:33:58.080 ⇒ 00:34:26.070 Uttam Kumaran: I was like, I just like, wish that we could just like reach that point. But it’s like I can like, take care of that sort of stuff as opposed. That’s why I’m I’m enjoying just working on some stuff internally, because I don’t need to really care so much about productionizing it when it comes to production, and people are paying for it to get the ensure that the structured outputs are the same at the time. Right?
195 00:34:26.150 ⇒ 00:34:35.419 Uttam Kumaran: That’s like these are 2, probably like, especially the guardrails, one is probably the first thing I saw. It’s a little bit about like, how do you set up guard rails
196 00:34:35.560 ⇒ 00:34:40.320 Uttam Kumaran: and expectations and tests that could be worth taking a look at.
197 00:34:40.489 ⇒ 00:34:43.850 Mohamed: We’re sure. No, thank you for sharing.
198 00:34:44.760 ⇒ 00:34:45.500 Mohamed: Okay.
199 00:34:47.320 ⇒ 00:34:48.469 that’s exciting.
200 00:34:48.550 ⇒ 00:34:53.720 Mohamed: Oh, yeah, there’s like so many tools, yeah, so many tools and whatnot?
201 00:34:54.210 ⇒ 00:34:58.430 Uttam Kumaran: So so yeah, go ahead. No, no, please.
202 00:34:59.140 ⇒ 00:35:12.999 Mohamed: Now, I was gonna ask what the structure like, what? So you’re working. Are you just working with Matt for like this one project, or what do you? What’s like the long-term structure like, what? What do you guys planning? I mean? So so far like. I came on board
203 00:35:13.030 ⇒ 00:35:15.169 Mohamed: like sort of like last week
204 00:35:15.240 ⇒ 00:35:20.570 Uttam Kumaran: on like a try period that should end on Friday.
205 00:35:20.870 ⇒ 00:35:32.819 Mohamed: assuming. So like we initially like, decided that we were going to like plan for the Mvp. They’re trying to push an Mvp. As fast as they can or like within like a few months
206 00:35:33.020 ⇒ 00:35:37.309 Mohamed: so I was like, yeah, sure
207 00:35:37.510 ⇒ 00:35:47.759 Mohamed: doesn’t like doesn’t seem like that difficult like with an Mvp. If we’re like looking at one or 2 like data source, like, if you have slack, if you have gmail, if you have zoom.
208 00:35:47.860 ⇒ 00:35:59.970 Mohamed: whatever select one or 2 that’ll be great, and it’s like utilizing the like seems pretty doable. And then like start like building it out like properly build functionality a lot more hired more people.
209 00:36:00.370 ⇒ 00:36:08.680 Mohamed: So I think that’s like, sort of like the plan. But then, yeah, a couple of days ago, they just like asked me to like, build a prototype. So I’m sick. Okay.
210 00:36:09.900 ⇒ 00:36:13.310 Mohamed: you’re like literally pressing me to like, build the connectors right now.
211 00:36:13.410 ⇒ 00:36:31.499 Uttam Kumaran: Yeah, I mean, I don’t know. I think the biggest thing you could probably push for is like, get a get. Get them to get a back end person. That’s probably what I’ll tell Matt to is like you need a back end person to not only handle hosting all that logic, but then also writing to connect your stuff to dude like that’s that’s like something that a lot of people have written before.
212 00:36:31.560 ⇒ 00:36:41.809 Mohamed: And I think the really, the key thing you should spend your month your time on is like energy is like on the language
213 00:36:41.850 ⇒ 00:36:54.580 Mohamed: that, like, it’s not that complicated. It doesn’t like. It’s not rocket science, like it’s literally been done like a thousand times over. But to like actually get it to work. It’s like too much like a bug, Lisa would say
214 00:36:54.890 ⇒ 00:37:03.160 Uttam Kumaran: no. But also you need to put your cycles on exactly if you have. Yeah, if you have, if you have a small team, let me focus on this other stuff here like this
215 00:37:03.170 ⇒ 00:37:04.670 Mohamed: whole other world.
216 00:37:05.170 ⇒ 00:37:22.349 Uttam Kumaran: Totally. No, I know. But your your iteration cycle need to be on like this sort of logic, not on like. Oh, I can’t communicate with like slack. You should be like someone is handling, taking that payload and routing it. And like, that’s
217 00:37:22.350 ⇒ 00:37:37.050 Mohamed: multi data tenant platform. They’re sort of like. That’s what they’re doing now. Sort of like building connectors.
218 00:37:37.060 ⇒ 00:37:49.489 Mohamed: putting them all into like postgres, giving you like a nice clean ui to like handle all of it, and that’s it. I was like, that’s perfect for me. It’s like, do. If you do that, that’ll be sublime. Then I can focus on the other stuff. But
219 00:37:49.490 ⇒ 00:38:10.069 Uttam Kumaran: yeah.
220 00:38:10.420 ⇒ 00:38:16.060 Mohamed: no, I mean, obviously, like, honestly, thank you so much for your time again. That’s a great meeting, you. But
221 00:38:16.110 ⇒ 00:38:21.530 Uttam Kumaran: more than anything. It was like messes like, bounce all these ideas and like, actually, like.
222 00:38:21.740 ⇒ 00:38:25.469 Mohamed: just make sure that, like, I’m not like sitting. Well.
223 00:38:25.570 ⇒ 00:38:42.270 Uttam Kumaran: yeah, yeah, yeah, no, it looks good. I mean, you got it all working, I think. Now again, it’s just like cycles. I’m like refining it and having it do different things, and so slacks so slacks slack. But I can actually send results back like you can put in little buttons and stuff like that. Consider that like, whether you can collect information.
224 00:38:42.320 ⇒ 00:38:50.489 Mohamed: you can fine tune and edit them to like, incorporate more down into it’s output. This is so so much fun stuff.
225 00:38:50.660 ⇒ 00:38:57.379 Mohamed: honestly. But yeah, I’m excited to see. And please, if you ever knew you weren’t hit me up.
226 00:38:57.730 ⇒ 00:38:59.950 Uttam Kumaran: Yeah, it’s all right. Well, I will definitely
227 00:39:00.080 ⇒ 00:39:06.389 Mohamed: alright! Alright! Sounds good, hope. See you soon alright. See you bye.