Meeting Title: Brainforge x Provenance: AI & Automation Date: 2025-06-12 Meeting participants: Anthony, Uttam Kumaran, Robert Tseng, Anton Romash
WEBVTT
1 00:01:07.860 ⇒ 00:01:08.810 Uttam Kumaran: Hey!
2 00:01:09.550 ⇒ 00:01:11.239 Anthony: Hey, Tom, how are you.
3 00:01:11.240 ⇒ 00:01:13.449 Uttam Kumaran: Hey? Good! How are you? Nice to meet you?
4 00:01:13.450 ⇒ 00:01:16.970 Anthony: Yeah. Nice to meet you as well. So how do you know, Craig.
5 00:01:17.730 ⇒ 00:01:26.210 Uttam Kumaran: Oh, how do I know, Craig? I got introduced to Craig from another founder friend who runs a
6 00:01:27.238 ⇒ 00:01:44.230 Uttam Kumaran: data company out of New York and then me and Craig have been talking for about a year. You know he’s in an interesting role at Crox, and then we’re helping a little. We’re helping him a bit with off the record and his program. And and also we’ve become friends. So we talk all the time about AI stuff.
7 00:01:44.746 ⇒ 00:02:00.230 Uttam Kumaran: We’re like on the ground doing a lot of implementation work. And so I’m pretty sober with like what’s actually working. And then he’s having these high level conversations and deciding stuff across and his own career. But Craig is like a very well like a network machine
8 00:02:00.460 ⇒ 00:02:12.125 Uttam Kumaran: like, I thought, okay for my business I’m like, Hey, we’re meeting a lot of people, but I’m some. I’ve never seen someone so regiment, you know. He’ll call me and say like, How can I help you? Tell me? Give you one thing to do for you, you know, and that’s
9 00:02:12.890 ⇒ 00:02:20.040 Uttam Kumaran: you know, that’s a great friend, and so he’s he’s been. He’s been really, really like, you know, great friend and great partner for our business. So.
10 00:02:20.460 ⇒ 00:02:29.509 Anthony: Yeah, it seems to parallel my experience. So, so yeah, I I can’t remember exactly how I was introduced to Craig anymore.
11 00:02:29.959 ⇒ 00:02:33.109 Uttam Kumaran: Yeah, see, we have the same challenge.
12 00:02:33.560 ⇒ 00:02:36.499 Anthony: But it’s been a very fruitful yeah.
13 00:02:37.450 ⇒ 00:02:38.029 Uttam Kumaran: You got it?
14 00:02:38.030 ⇒ 00:02:38.999 Uttam Kumaran: Yeah, that’s awesome.
15 00:02:39.000 ⇒ 00:02:40.390 Anthony: Hey! Rob! So.
16 00:02:40.390 ⇒ 00:02:41.300 Uttam Kumaran: Hey!
17 00:02:41.300 ⇒ 00:02:49.139 Anthony: So yeah, I. I helped him with some intros to some hosts, for you know, off the record. And he’s just
18 00:02:49.850 ⇒ 00:03:00.190 Anthony: helping us on the side to how to think about AI. And so that’s how the the introduction was made. So excited to to be chatting nice to meet you, Robert.
19 00:03:00.190 ⇒ 00:03:01.609 Robert Tseng: Good to meet you, Anthony.
20 00:03:04.270 ⇒ 00:03:09.148 Uttam Kumaran: Nice to meet you guys. Yeah, I love the I love the the background like I need.
21 00:03:09.590 ⇒ 00:03:19.539 Uttam Kumaran: you know I’ve been. Look I have in my Amazon card of a bunch of like Lakers stuff that I want to put up like, and I have my like dogs, Lakers bandana here.
22 00:03:19.540 ⇒ 00:03:20.170 Anton Romash: Nice.
23 00:03:20.499 ⇒ 00:03:26.100 Uttam Kumaran: Need to put. I need to put on some like some basketball stuff up to remind me.
24 00:03:26.430 ⇒ 00:03:34.820 Anton Romash: Listen. I I won’t tell you how long it’s sat around my office, you know, before I actually took the time to hang it up.
25 00:03:35.780 ⇒ 00:03:37.119 Uttam Kumaran: That’s great.
26 00:03:37.270 ⇒ 00:03:50.019 Uttam Kumaran: Well, well, cool. It’s awesome to get connected. Guys. So tell me how we can be helpful or I know I know I got a little bit of info from Craig. But not too much. So yeah. Feel free.
27 00:03:50.020 ⇒ 00:03:59.721 Anthony: Yeah. The the long story short is when we’re chatting with Craig. You know, one of the things we’re trying to suss out is,
28 00:04:00.300 ⇒ 00:04:22.140 Anthony: you know whether in certain areas of our business. So maybe I should take a step back and just tell you what we do first, st so that there’s some context. So we’re, we’re a growth stage investor and consumer brands. And so typically that means we’re getting involved when the companies are around 50 ish 1 million or revenue is is the average. So
29 00:04:23.341 ⇒ 00:04:44.428 Anthony: we have a very data driven front end on diligence. So but it’s all 1st party data that we’re getting from companies. So we’re ingesting their transaction log customer file doing predictive analytics, layering on a bunch of other demographic and psychographic data on top. We won’t bore you those details, but
30 00:04:45.210 ⇒ 00:04:49.700 Uttam Kumaran: Not boring, that’s what we that’s all we do. So yeah.
31 00:04:49.700 ⇒ 00:04:53.239 Anthony: So we we do a bunch of this stuff. And so.
32 00:04:53.390 ⇒ 00:05:06.731 Anthony: you know, for every company, there’s lots of things that we’re just gonna do over and over and over again. And so in my mind, those things are candidates for automating, and so
33 00:05:07.830 ⇒ 00:05:15.890 Anthony: what what I’m trying to figure out is, you know, are there efficient ways to, you know, use agents
34 00:05:16.800 ⇒ 00:05:24.980 Anthony: for for what we’re doing on? Maybe the more I’d say qualitative side of of research.
35 00:05:25.420 ⇒ 00:05:29.220 Anthony: That’s, you know, research based
36 00:05:29.390 ⇒ 00:05:36.190 Anthony: crawling the web for information on various brands, etc. So that’s the backdrop of
37 00:05:36.870 ⇒ 00:05:42.899 Anthony: of what we’re trying to think through. You know, we can have some conversations around some specific tasks, but.
38 00:05:43.190 ⇒ 00:05:43.960 Uttam Kumaran: Cool.
39 00:05:43.960 ⇒ 00:05:48.980 Anthony: But yeah, figured I’d start there as general context. So with that.
40 00:05:49.170 ⇒ 00:05:51.430 Anthony: I don’t even know if that’s in your guys.
41 00:05:51.430 ⇒ 00:05:57.370 Uttam Kumaran: No, that’s helpful, let me tell. Let me tell you a little bit about us. I think we actually have a lot more overlap
42 00:05:57.795 ⇒ 00:06:20.850 Uttam Kumaran: then I then I initially read. But I checked out the website. So we we’ve actually done a lot of work with consumer brands. You know, I’ve I’ve done some work at athletic greens. We we have a bunch of clients that are in that sort of 20 to 100 million dollars spot where we we do a lot of data analytics, predictive analytics, product measurement, profitability measurement. So that’s really.
43 00:06:21.250 ⇒ 00:06:46.169 Uttam Kumaran: I don’t know. We’ve worked with a number of clients in that space so super super familiar with what those brands do, what they care about and so one, I think that’s that’s really helpful. Second on the AI side. This is really something that to be frank, we started the business about a year and a half ago, and we were. We’ve been using AI to try to automate our own business. And our backgrounds have been as operators and data. So we’ve run data teams run analytics functions
44 00:06:46.170 ⇒ 00:07:06.509 Uttam Kumaran: work directly with executives. And so we were using AI just to like automate our business. So our sales process our execution process. How we qualify leads. And then also how we manage and and execute, you know, on our data projects. And in that process, you know, we built out an AI team that was serving us internally. And then.
45 00:07:06.510 ⇒ 00:07:24.769 Uttam Kumaran: of course, you have the brain blast moment where it’s like, this is actually kind of hard to do. I’ve been using AI, you know, just since Gpt. 3.5 came out, but moving from that to like orchestrating a workflow, having the right data in and actually having something that’s not like a gimmicky thing where you like download some sas tool, and it’s like
46 00:07:25.120 ⇒ 00:07:39.070 Uttam Kumaran: on like AI. Sdr, it’s like, not that easy at all. And so we’ve now come to the come to some really great conclusions internally, and we actually now do this for customers as well. So both sides of our businesses are one. We help
47 00:07:39.340 ⇒ 00:07:49.719 Uttam Kumaran: ambitious growing businesses. Basically shore up and modernize their data analytics. And then we try to make that actionable through AI,
48 00:07:50.910 ⇒ 00:08:12.219 Uttam Kumaran: so your guys process is actually really perfect, because I think you probably have a very methodical process of the stuff you collect. I think what you mentioned is, there’s maybe qualitative information that you’re missing. There’s also probably some stuff, even in the quantitative part of that process that you could automate. You know parts of it. And then there’s basically like having AI also flagged
49 00:08:12.310 ⇒ 00:08:19.410 Uttam Kumaran: things in that process that you should look at versus others. So there’s probably a couple of areas for us to go down, I think just to go on your
50 00:08:19.460 ⇒ 00:08:48.299 Uttam Kumaran: sort of qualitative example. Certainly, like, I think there’s a lot with like using Openai search, perplexity, search to gather information, to come to conclusions. We build things that are both like copilots so that you can chat with over a bunch of documents. We also build things that go and do stuff like, hey, take an account like for, for in our use case we do a lot of account and technology based marketing where you want to take an account or technology, go find a bunch of information, find clients that are using that
51 00:08:48.360 ⇒ 00:08:58.410 Uttam Kumaran: you have agents that have multiple tools that go orchestrate, a bunch of steps come back to me or Robert, or whoever with some sort of output that we then can discern and take action on
52 00:08:59.930 ⇒ 00:09:01.040 Uttam Kumaran: So yeah, I mean.
53 00:09:01.220 ⇒ 00:09:21.370 Uttam Kumaran: I also just were vomited a lot as well. So tell me, like, if it’s if you think it’s be more helpful to maybe even just walk through like a quick example. Or you know, for our from our side, the way we try to start these is really isolating, like a very specific part of the prompt. Because, as you know, like.
54 00:09:21.500 ⇒ 00:09:34.439 Uttam Kumaran: there’s just a lot of avenues to take this that way, you know, even in just explaining to you what is possible, I can describe specific tools, specific, like examples, and even follow up with like
55 00:09:34.640 ⇒ 00:09:38.449 Uttam Kumaran: couple of ways that you guys can go play around with something, you know.
56 00:09:38.940 ⇒ 00:09:43.532 Anthony: Yeah, I I’d love to see an example of how you guys are using it today.
57 00:09:44.080 ⇒ 00:09:44.760 Anthony: go from there.
58 00:09:44.760 ⇒ 00:09:49.822 Uttam Kumaran: Yeah, maybe I can. Even, I may even just pull up
59 00:09:51.040 ⇒ 00:09:55.165 Uttam Kumaran: like one of our like lead research things that’s in our slack. Robert.
60 00:09:56.930 ⇒ 00:10:00.029 Uttam Kumaran: let me let me find one that’s that’s good.
61 00:10:09.000 ⇒ 00:10:18.416 Uttam Kumaran: So here is an example and sorry no prep here. So I’m just gonna just pull up our slack to show you an example of something we’re doing. So we
62 00:10:18.880 ⇒ 00:10:20.990 Uttam Kumaran: of course, just like everybody. We
63 00:10:21.090 ⇒ 00:10:39.859 Uttam Kumaran: go after accounts, and we go after people in those accounts and try to put together a story on why, you know, we think we can help them. And so this is an example of something that started with a Hey, Robert and I, when we meet a company, or get introduced to a company, or get introduced to a person or someone’s like, Hey, I have a I have someone I can intro you to.
64 00:10:40.090 ⇒ 00:11:00.220 Uttam Kumaran: We then have to spend anywhere from 15 to an hour to just like, Okay, what’s the story here? Right? And so we in probably a few days. And we this we I don’t know. This has been out on our internal in our company since like December. But we built a little like lead researcher. So what this does is we send it a company?
65 00:11:00.606 ⇒ 00:11:11.040 Uttam Kumaran: We then send it like a person in the company. And then it goes and does. A bunch of research for us finds like recent funding news, relevant services qualification.
66 00:11:11.530 ⇒ 00:11:14.420 Uttam Kumaran: But there’s a couple of steps beyond, just like.
67 00:11:14.900 ⇒ 00:11:42.060 Uttam Kumaran: okay, if I was to throw this in Chat Gpt, what’s the difference? One is, the researcher has understanding of what we do and what we sell, and to be specific, it has understanding of all of our case studies. It has a lot of prompt engineering on the exact services we offer. It also has information about the way we service and the tools we use. So when it comes back to us, it tells us the stuff that we’re interested in.
68 00:11:42.384 ⇒ 00:12:02.810 Uttam Kumaran: So like, what are the pain points as it relates to data or AI, right? So the need for secure and efficient communication tools, workflow automation, patient engagement. These are all like very accurate, very top of mind for these types of companies. And then it does some recommendation on, hey? Here are relevant services in our service, offering that we could offer
69 00:12:02.910 ⇒ 00:12:07.529 Uttam Kumaran: it. Then says, Okay, let’s do some qualification. Okay, they raise this
70 00:12:07.780 ⇒ 00:12:18.390 Uttam Kumaran: that there. Here’s a decision, timeline. Here’s some competitive advantage, and we’re not, you know. Of course, we’re not taking this, and like robotically copying this and spewing this out. But for us
71 00:12:18.808 ⇒ 00:12:29.420 Uttam Kumaran: like this saved 30 min for me to go from. Are they good, or are they bad? Can I draft an email real quick and get something out so that that person can make an intro in the next 10 min.
72 00:12:30.217 ⇒ 00:12:33.072 Uttam Kumaran: And so this is like one example.
73 00:12:33.900 ⇒ 00:12:45.021 Uttam Kumaran: we’re doing. We’re to think about more like of this like search based thing. We. We also use a lot of enrichment related sources. So if I was to show
74 00:12:45.490 ⇒ 00:12:47.800 Uttam Kumaran: one more example. Here.
75 00:12:48.265 ⇒ 00:13:10.329 Uttam Kumaran: give me a second. Let me just pull it up. We do a couple of sort of. We run a couple of different marketing place. So we run like account based marketing like you just saw. We also do like technology, and so one of the things that’s really interesting to us is like, can we go find all of the folks using a specific technology that fits our Icp so that we can then craft a narrative?
76 00:13:10.837 ⇒ 00:13:13.373 Uttam Kumaran: Towards that. So let me just pull up
77 00:13:14.800 ⇒ 00:13:20.839 Uttam Kumaran: this sort of workbook. And if are you guys familiar with the company called Clay. By the way, if you’ve heard of Clay.
78 00:13:20.840 ⇒ 00:13:21.500 Anthony: Yep.
79 00:13:21.813 ⇒ 00:13:36.859 Uttam Kumaran: They’re like a big go to market automation platform that just like they just, you know, more, raised a bunch of money recently. But for example, we do a lot of work with with like amplitude. And so one of the things that we wanted we
80 00:13:36.990 ⇒ 00:14:02.760 Uttam Kumaran: I I tasked our team to do is say, Hey, Robert and I are interested in going and targeting folks that are using amplitude. But that’s not just like you know, going into like that’s a comes from a couple of different areas. One is like, I want us to go to Amplitude’s website, scrape any Logos, and then go use those Logos to go find those companies. Second is, I want to go. Use a couple of different services to find any job posting
81 00:14:02.810 ⇒ 00:14:18.830 Uttam Kumaran: from any company that references amplitude, as like a part of the requirements which indicates their user. Once we then have that list, we then need to go through some qualification process. So we use clay heavily because it’s very workbook based. So we can go from a company
82 00:14:18.880 ⇒ 00:14:35.690 Uttam Kumaran: to then dissecting, like its name, finding a bunch of adjacent information. And then we then send this to our account based marketing play which finds our Icps. It finds like what are potential case studies that could be relevant to them, and then goes even further and helps us draft that messaging.
83 00:14:36.088 ⇒ 00:14:39.289 Uttam Kumaran: But this is where I think we’re a little bit more opinionated is like.
84 00:14:39.390 ⇒ 00:15:06.540 Uttam Kumaran: there’s not an AI agent that’s like just doing this automatically. I don’t really think that’s like worth handing off. We sell, we sell consulting services. People want to talk to us, and we want to also talk to them and share and find out for a fit. So this all just kind of helps us speed up. Our process improves time between, follow ups times to close deals, but also helps us disqualify and target people better.
85 00:15:07.830 ⇒ 00:15:13.102 Uttam Kumaran: so those are a couple of examples. I you know we have a handful of more. I can share.
86 00:15:13.380 ⇒ 00:15:15.319 Anthony: If you don’t mind my asking. I know you said
87 00:15:16.110 ⇒ 00:15:22.460 Anthony: people who work with amplitude. So how are you getting that list, in the 1st place, to to feed.
88 00:15:22.830 ⇒ 00:15:24.787 Uttam Kumaran: Yeah, so there’s
89 00:15:26.000 ⇒ 00:15:28.050 Uttam Kumaran: There’s 2 kind of
90 00:15:28.210 ⇒ 00:15:39.196 Uttam Kumaran: big companies in this space. One is this tool called built with, this is very famous. They’re an enrichment provider that basically, there’s another tool Rob, or I forgot what the name is.
91 00:15:40.510 ⇒ 00:16:01.539 Uttam Kumaran: there’s another enrichment source that we use that actually goes and finds the job postings. But on top of that we also run several different Google search parameters pull in all of that information like, scrape. Those sites pull that in. And we run some scrapers on top of the technology site to find these.
92 00:16:01.650 ⇒ 00:16:14.639 Uttam Kumaran: So you know, you may get like 40% from here 40% from another source. It’s kind of like the orchestration that matters. And then, finally, what’s really important for us is that this ends up in slack.
93 00:16:14.880 ⇒ 00:16:19.430 Uttam Kumaran: because that’s where we work right, and we work in slack and notion. So for us, it’s important that
94 00:16:19.600 ⇒ 00:16:30.733 Uttam Kumaran: it’s not so manual for us to have to go to Clay and do that. It’s something that where I can send an account or a technology in things run, and then where we get back some action to take.
95 00:16:31.620 ⇒ 00:16:39.350 Uttam Kumaran: So that’s a couple of things. I mean, I I would, you know, love to hear more about the specific like diligence
96 00:16:39.470 ⇒ 00:16:45.929 Uttam Kumaran: related, use case or the things that you’re searching. I could share some really fantastic search.
97 00:16:46.320 ⇒ 00:16:51.100 Uttam Kumaran: and, like search related like deep research and search related Apis that we use for
98 00:16:51.710 ⇒ 00:16:54.240 Uttam Kumaran: for some of that more qualitative piece, as well.
99 00:16:55.290 ⇒ 00:17:10.650 Anthony: Yeah. So I’d say, as we listed out the the things that we we wanted to potentially automate, I I don’t even know if this is is worth doing. But one topic was, you know, could we
100 00:17:10.930 ⇒ 00:17:23.945 Anthony: go into a brand on the diligent side? Could we go into a brand’s website? And basically scrape all the reviews and have it automatically dumped into
101 00:17:24.819 ⇒ 00:17:31.839 Anthony: a database that we could, you know, create a word cloud with or analyze in some other way.
102 00:17:31.840 ⇒ 00:17:35.389 Uttam Kumaran: From like trustpilot or or from sure. Okay.
103 00:17:35.390 ⇒ 00:17:38.130 Anton Romash: Pilot in their own website, maybe Amazon.
104 00:17:38.390 ⇒ 00:17:39.170 Uttam Kumaran: Cool.
105 00:17:41.020 ⇒ 00:17:51.199 Anthony: So that that was one thing that we thought we could do on the diligence side. What? What you just showed was something that we talked about hypothetically on the sourcing side.
106 00:17:51.200 ⇒ 00:17:52.720 Uttam Kumaran: Sourcing. Side, yeah.
107 00:17:52.840 ⇒ 00:17:53.574 Anthony: Yeah.
108 00:17:55.690 ⇒ 00:17:56.190 Anthony: Good.
109 00:17:56.190 ⇒ 00:18:00.249 Uttam Kumaran: Sort of helpful to understand. You know what I tell our clients is like.
110 00:18:00.610 ⇒ 00:18:05.990 Uttam Kumaran: It’s sort of where you have, like the same 4 quadrants of like urgent and important. You find out what is the
111 00:18:06.290 ⇒ 00:18:21.359 Uttam Kumaran: I can tell you what’s easy and hard, but then I think what’s important for us to hear is what what is worth doing like, what is something that? Okay? If we had in like the next 2 weeks would really land us like 20 to 30% more time.
112 00:18:21.500 ⇒ 00:18:25.710 Uttam Kumaran: And then there’s gonna be some stuff where I’m like, okay, that’s gonna take, like, probably like
113 00:18:26.120 ⇒ 00:18:28.869 Uttam Kumaran: 1, 2 months to to execute on.
114 00:18:29.050 ⇒ 00:18:34.510 Uttam Kumaran: But even anything that’s like that sounds like scraping something moving into a spreadsheet.
115 00:18:34.860 ⇒ 00:18:45.279 Uttam Kumaran: Very table stakes, the really, the complicated pieces on that is, if you’re going to a site where you have to authenticate and log in. There’s a little bit complication. Or if the data is not so structured.
116 00:18:45.658 ⇒ 00:18:51.779 Uttam Kumaran: then you have to have like, take a bunch of text, have a prompt synthesize that, do something. But I think also
117 00:18:51.930 ⇒ 00:18:57.579 Uttam Kumaran: part of why we got into this business is most of this is actually like data engineering.
118 00:18:57.650 ⇒ 00:19:23.209 Uttam Kumaran: It’s like moving data around the Llm and like shoving context into it. It’s sort of the way I describe it is like when you have, like your perfect meeting transcript, you have your like Crm data and you paste it all of the chat gpt with a great prompt. You’re getting a great output, but you need to do that every time. Multiple times a day. Everybody your team who nobody has any idea how to use that needs to be able to do it. And that’s like a data orchestration
119 00:19:23.670 ⇒ 00:19:29.449 Uttam Kumaran: problem less you that the proof of concept is exactly that moment.
120 00:19:29.580 ⇒ 00:19:40.779 Uttam Kumaran: but enabling that in your slack Workspace, or over your email, or on some cadence ad hoc, and then having it all work and deliver something back to you. That’s the
121 00:19:41.390 ⇒ 00:19:45.059 Uttam Kumaran: where building these agents or agentic workflows comes in.
122 00:19:46.670 ⇒ 00:19:47.320 Anton Romash: Yeah.
123 00:19:47.943 ⇒ 00:19:57.909 Anthony: That makes sense. So with your the clients that you you are working with, how? How do you tend to to work with them. What’s the work structure.
124 00:19:58.550 ⇒ 00:20:13.340 Uttam Kumaran: Yeah. So for all of our clients, we, we 1st try to do something more advisory and just arrive at like a proof of concept so if we can agree on something that at least we could time box to one or 2 weeks where.
125 00:20:13.710 ⇒ 00:20:27.270 Uttam Kumaran: you know, we sort of work together, you get a sense of how we plan and arrive at like milestones in a project, and then we deliver something for you that works that you could sort of see all within a 1 to 2 week
126 00:20:27.480 ⇒ 00:20:31.019 Uttam Kumaran: span. That’s where we like to start. I think it gives.
127 00:20:31.110 ⇒ 00:20:58.980 Uttam Kumaran: you know, our clients a really good sense of like what it’s like to work with us? But also gives us a sense of like, is there actually a very clip? Can we define a clear problem at the end of that sort of proof of concept you’re left with actually, like a couple of artifacts. You’re left with one ideally like what an Mvp. And then what like the next couple of versions, and like a timeline second is we. We do a lot of documentation. So you’ll be left with, like, what are the tools we use? How would it get built? So I would say.
128 00:20:59.250 ⇒ 00:21:23.609 Uttam Kumaran: we try not to like gatekeep any of that I actually try to. I want to show you like, what what’s complicated, what’s easy. And I want to really share that. And then at that point. What what our goal is that you decide to work with us on the broader implementation. But you also have those assets that you can go shop around. So for us the most important thing is, if we can isolate one use case.
129 00:21:23.990 ⇒ 00:21:29.590 Uttam Kumaran: then I could break down what a proof of concept could look like. And then what an Mvp. Could look like, and
130 00:21:29.720 ⇒ 00:21:36.850 Uttam Kumaran: ideally, we would aim to do some something that’s around a week or 2 weeks where we can structure some milestone.
131 00:21:37.725 ⇒ 00:21:42.420 Uttam Kumaran: Whether it is the qualitative part of the diligence, whether it’s a sourcing side.
132 00:21:42.540 ⇒ 00:21:59.046 Uttam Kumaran: I think, for us. What’s important is that you pick something that’s that is really, really important. And that is really taking a lot of time. And ideally, something that is you’re like, I don’t know whether this can be automated like something challenging for us to really take on show.
133 00:21:59.370 ⇒ 00:22:01.029 Robert Tseng: Jump in on that. Like, I think.
134 00:22:01.230 ⇒ 00:22:16.139 Robert Tseng: even just, I just want to emphasize that like, yeah, we’re working toward the proof of concept. But we fully expected to go to production. I do think that there are a lot of folks out there that are. I mean, you could probably vibe code a proof of concept for yourself. But you have no idea how to bring it to production necessarily. And like, I think.
135 00:22:16.490 ⇒ 00:22:40.370 Robert Tseng: really finding a problem that is actually going to hit production, I think, is what, in our best interest, like your time didn’t mention this. But like with Vitaco, for example, I think that’s a great example of like, hey? Through like a discovery that we did with them. Turns out we felt like we could go and figure out their inventory stock outs across the entire digital marketplace faster than their internal teams. So we just built an AI agent for that.
136 00:22:40.370 ⇒ 00:22:49.080 Robert Tseng: It’s in production. And that was like, you know, that was a huge win, and that was easy to deploy, but also kind of like showed that we exceeded.
137 00:22:49.080 ⇒ 00:23:11.139 Robert Tseng: you know, their their existing capabilities. So like, I think, you know, be able to deploy something at Enterprise Scale very quickly. It’s not far off, like we definitely have the capability to do that. And obviously, this is more for internal tooling for you. But I think just wanna stress the fact that like, yeah, like this, we we fully expect it to be used, and that’s part of our metric of success.
138 00:23:12.800 ⇒ 00:23:13.666 Anthony: Got it.
139 00:23:14.851 ⇒ 00:23:18.340 Anthony: So hypothetical if we had a
140 00:23:18.892 ⇒ 00:23:27.337 Anthony: you know, kind of a client profiling tool like you guys just created we were using it for deal sourcing
141 00:23:27.850 ⇒ 00:23:31.709 Anthony: instead of, you know, tapping into built with.
142 00:23:32.320 ⇒ 00:23:35.950 Anthony: let’s say, we’re tapping into a a different data set.
143 00:23:37.940 ⇒ 00:23:38.730 Anthony: How
144 00:23:39.080 ⇒ 00:23:46.659 Anthony: how difficult. Do you think that would be for for us like repurposing? Obviously the the workflows would be very similar.
145 00:23:47.050 ⇒ 00:23:53.380 Anthony: What you’re instructing, you know the agent to do along the way is a little bit different and
146 00:23:55.210 ⇒ 00:23:58.809 Anthony: but I think the flow of information will probably be fairly similar.
147 00:23:59.420 ⇒ 00:24:02.809 Uttam Kumaran: Yeah. So what I think one is like, we’ve
148 00:24:03.020 ⇒ 00:24:12.499 Uttam Kumaran: as an AI company, and we’ve used a lot of AI tools. There’s like a hundred coming out every week. There’s a lot that is like complete vaporware. And there’s a lot that works so
149 00:24:12.600 ⇒ 00:24:22.189 Uttam Kumaran: clay and a couple of these tools that we typically build on are things that are gonna exist for a while. And are actually, you have the ability to go in and debug even as a non technical user.
150 00:24:22.514 ⇒ 00:24:45.760 Uttam Kumaran: So that’s 1 thing that’s really important to us when we pick tools. Second piece is, we just would need to sit with you and understand. Okay, what is, what are, what is the Icp, what are data points that you have or don’t have today? That would be relevant and like, what does the output look like? Is it a checkmark like this is a quality someone that’s qualified? Is it like a summary? Where does that live? So we would have to go through that discovery
151 00:24:45.930 ⇒ 00:25:04.759 Uttam Kumaran: sort of process and get those requirements. But I think something on the sourcing side is something we would basically put together in a week or 2. And you would, I think you would actually be able to basically see and touch and play around with something that’s working in that timeframe. Pretty confidently.
152 00:25:05.500 ⇒ 00:25:06.260 Anthony: Okay.
153 00:25:06.750 ⇒ 00:25:21.950 Uttam Kumaran: Of course, like the what you get in the speed to deployment. It’ll take time to make sure it works, because you’ll get some. You’ll get an output, and then it’ll be tuning and making sure that like, okay, it’s actually saving me time. It’s not just another thing I have to like
154 00:25:22.500 ⇒ 00:25:27.209 Uttam Kumaran: manage, you know, and that’s where I think we’ve pushed our internal team.
155 00:25:27.340 ⇒ 00:25:37.859 Uttam Kumaran: And part of you know, I think our advantage is like we’re using a lot of this stuff internally, and I know that it’s like it can’t just be another tool that I have to go to another ui and do stuff in like that’s like
156 00:25:38.550 ⇒ 00:25:44.540 Uttam Kumaran: that. That’s like dead on arrival. So for us it has to live where we do work, and it has to
157 00:25:44.650 ⇒ 00:26:00.030 Uttam Kumaran: take has to really confidently take time off of my plate or help speed up activities, things that I would probably have handed off to a coordinator or Ea, or like a junior person. Those are the types of tasks that reserves us to actually like.
158 00:26:00.260 ⇒ 00:26:04.259 Uttam Kumaran: Do the thing we’re good at, which is like, make the distinction, or add the
159 00:26:04.410 ⇒ 00:26:08.130 Uttam Kumaran: the little magic on top, and then, you know, do something with. So
160 00:26:08.330 ⇒ 00:26:13.109 Uttam Kumaran: if that seems like a good use case, then that’s what I would. That’s probably what I would.
161 00:26:13.290 ⇒ 00:26:26.579 Uttam Kumaran: What aim to propose. And on our side, we would just structure like a small paid proof of concept phase that we would aim for 2 weeks. I can send the details out for this, and if we can agree on
162 00:26:26.840 ⇒ 00:26:30.340 Uttam Kumaran: that scope, and I think that’s a good place to start.
163 00:26:31.180 ⇒ 00:26:35.249 Anthony: Yeah, yeah, that sounds good. Yeah. We’re
164 00:26:35.970 ⇒ 00:26:39.709 Anthony: we’re deeply curious to try to. You know.
165 00:26:39.830 ⇒ 00:26:43.979 Anthony: automate some stuff here, I I think for for us. It’s just about
166 00:26:45.220 ⇒ 00:26:49.877 Anthony: the cost benefit analysis on any piece of it has to make sense.
167 00:26:50.210 ⇒ 00:26:50.810 Uttam Kumaran: Yes.
168 00:26:51.400 ⇒ 00:26:54.289 Anthony: Obviously cause for for us. It’s it’s
169 00:26:55.240 ⇒ 00:27:03.479 Anthony: we. We don’t expect that it’s going to turn up a lot of opportunities that were otherwise missing. It’s more about the efficiency.
170 00:27:05.590 ⇒ 00:27:06.280 Uttam Kumaran: I agree.
171 00:27:06.280 ⇒ 00:27:13.900 Anthony: You know, it’s sussing through vetting through stuff more rapidly is is the goal.
172 00:27:14.290 ⇒ 00:27:21.309 Uttam Kumaran: Do you think it’s more about like we? There’s other information out there that we you may not have considered? Or is it
173 00:27:21.430 ⇒ 00:27:29.580 Uttam Kumaran: it like between that and speed. That’s probably where I see, but, like it may be like a 70, 31 way or the other.
174 00:27:30.610 ⇒ 00:27:33.880 Anton Romash: Honestly, I think it’s yeah. It’s it’s it’s speed.
175 00:27:34.480 ⇒ 00:27:47.259 Anton Romash: If you can think of it as speed, or you can think of it as bandwidth. Right? Like we, we. There’s a benefit for us to keep our team fairly lean and just not throw bodies at the problem. There’s all sorts of benefits. Right? And so the idea is like.
176 00:27:47.630 ⇒ 00:27:53.676 Anton Romash: can we take a lot of the busy work off of, you know, everyone’s plate?
177 00:27:54.790 ⇒ 00:28:06.300 Anton Romash: cause it’s truly just like pulling information together from disparate sources. A lot of which, like you said, are that you just have to log in right and and like sometimes it’s.
178 00:28:06.550 ⇒ 00:28:20.340 Anton Romash: you know. Sometimes the information is like Bs right, because, like the quality is not always consistent across the board. But between all the various sources you can kind of triangulate right? So when any one of us like sits down for an hour and just
179 00:28:20.750 ⇒ 00:28:41.549 Anton Romash: goes through a Brand’s Instagram feed and looks it up on pitch book and looks it up on a credit card panel and then looks it up on, you know, and it’s like by the end of it like we’ve pieced together a bunch of screenshots and slides that tell a great story in 10 slides. We’d love to be able to just like automate at least that piece of it.
180 00:28:44.100 ⇒ 00:29:06.230 Uttam Kumaran: I also think it’s even as you accumulate this information. You know, for us, we have a very similar problem. We we work with a lot of clients. But we have a we have a lot of information that gets generated through that entire process from the contracts to the meetings we have on zoom to the conversations we have internally, externally, in our slack channels about them. To the code we get pushed
181 00:29:06.370 ⇒ 00:29:17.309 Uttam Kumaran: to the documents that we write and what we do is all of that, I said, needs to get put behind an agent. Cause I I want our questions like, Hey, where is this document.
182 00:29:17.530 ⇒ 00:29:29.500 Uttam Kumaran: What did we talk about last week like, did we? What like? What did we drop anything that I forgot to follow up on? Those are things that like? Yes, on one client, sure. But we have Pm’s that we want to stretch
183 00:29:29.590 ⇒ 00:29:50.189 Uttam Kumaran: like not only just a 3 clients, but 5, but 7 right? And we. I want to push back and see. I want to push back on. Okay, we have to hire, because that’s what consultancies do like. I want to see whether we can stretch and hand off the 20 to 30% which we know is so painful work and typically gets dropped because it is tough to go back and comb through. Did I miss anything this week like
184 00:29:50.190 ⇒ 00:30:17.049 Uttam Kumaran: who wants? No, no one wants to do that. The other thing is like, hey? If I want to ask a genuine question about hey? Given all these documents like, what opportunities should I pitch our clients on other like data initiatives, they should take on. That’s something that look. If I took all of that and shoved into Chat Gbt, it probably wouldn’t fit in the context. And like that’s there is challenges there. And and that’s also something for me. It’s not just about me. Getting enabled is everybody on our team having access to that.
185 00:30:17.190 ⇒ 00:30:29.460 Uttam Kumaran: because, that’s how this scales. And so for us, we think about the exact same problem which is like all these assets get created, and a lot of them like you read it once, and you’re like, okay, I wish I could just bookmark it for later.
186 00:30:29.510 ⇒ 00:30:34.969 Uttam Kumaran: But then you want it to be able to have something more dynamic to chat over the life cycle, of whether it’s a
187 00:30:34.990 ⇒ 00:31:01.350 Uttam Kumaran: portfolio company or or a lead or something, someone that’s going through diligence. And that can evolve into many things, you know, over time. But like that is a that is what we do on the data side, which is pull structured, unstructured documents together we do rag on top of that, retrieve the right things for the right time. And then allow the agent to to answer your question. So it’s not just like pulling from like
188 00:31:01.480 ⇒ 00:31:05.080 Uttam Kumaran: the 5 links it just found. It’s pulling from all of your proprietary
189 00:31:05.380 ⇒ 00:31:11.849 Uttam Kumaran: stuff. And that’s the edge that we’re seeing. But that takes time, you know it’s not. It’s it. It takes time to build.
190 00:31:13.980 ⇒ 00:31:14.480 Anton Romash: Yeah.
191 00:31:14.480 ⇒ 00:31:15.000 Anthony: So.
192 00:31:16.060 ⇒ 00:31:24.359 Anton Romash: Yeah, I think for us, we might just need, like a quick internal huddle, to figure out the best
193 00:31:24.670 ⇒ 00:31:28.210 Anton Romash: kind of thing to focus on for this Mvp stage.
194 00:31:30.550 ⇒ 00:31:40.319 Anton Romash: whether it’s more on the sourcing side or more on like the information gathering side, or whether they’re kind of just, you know part one and part 2 of the same exercise.
195 00:31:42.570 ⇒ 00:31:58.570 Uttam Kumaran: Yeah. And let me leave you with even like one more thing, maybe to help narrow that down. One is like, you want to have a really good understanding of what the end state you’re going for is because part of our process is we we do what’s called evaluations. So we run evals on everything where
196 00:31:58.750 ⇒ 00:32:25.239 Uttam Kumaran: we make sure. We want to make sure that when our AI agent responds to you, we score its response, based on some what we call a golden data set where we have structured question answers that have been approved like this question gets asked. The answer should be close or near this, otherwise you’ll anecdotally be like, Oh, cool, it’s magic, it’s like working. But then you have no sort of like an actual qualitative approach to measurement. So
197 00:32:25.440 ⇒ 00:32:31.400 Uttam Kumaran: as part of this, we will set that up for you where we typically do 25 to 50 like question, answer pairs
198 00:32:31.560 ⇒ 00:32:45.059 Uttam Kumaran: that you have, we have to actually, methodically go down and be like. This is what I would expect it to do that way. We have to judge whether it’s working, and then we want to see those scores go up over time. And so I would think about problems where
199 00:32:45.360 ⇒ 00:33:01.490 Uttam Kumaran: you know that that is possible where? Okay, I I. We’ve done this enough times. I have enough, and that doesn’t even mean you have to have it in your brain. But you have access to all the past times we’ve done this, whether those are emails or anything where we can start to break that down.
200 00:33:02.039 ⇒ 00:33:14.320 Uttam Kumaran: And you have a clear expectation of like what the output looks like. The cleaner that is, the easier this will be. And then there are a lot of use cases where it’s not that clean where we’ll have to discuss like what is
201 00:33:14.510 ⇒ 00:33:29.429 Uttam Kumaran: like, what do we expect out of like a lead, like a like a lead report, or something like that but the cleaner your expectations could be at this point. The more easy it is for us to to score that that proof of concept.
202 00:33:31.010 ⇒ 00:33:31.680 Anton Romash: Yup!
203 00:33:33.340 ⇒ 00:33:34.106 Anthony: Got it.
204 00:33:35.160 ⇒ 00:33:38.210 Anthony: Yeah, that makes sense cool.
205 00:33:39.210 ⇒ 00:33:45.909 Anthony: Okay. So I guess we’ll wait to see something from you guys. And then we can in the meantime, think about
206 00:33:46.410 ⇒ 00:33:49.809 Anthony: where we want to point the weapon on on this project.
207 00:33:50.340 ⇒ 00:34:00.710 Uttam Kumaran: Cool. Yeah, I’ll send a couple of things your way, and even some of the stuff that I shared today and a few others I’ll share. They’re like open demos that we have, that you can poke around at
208 00:34:01.197 ⇒ 00:34:04.229 Uttam Kumaran: and then, yeah, like any questions. In the meantime.
209 00:34:04.650 ⇒ 00:34:13.610 Uttam Kumaran: feel free. This is like all we’re all we’re doing every day. So it’s really interesting stuff. It’s crazy that it’s like a.
210 00:34:14.040 ⇒ 00:34:36.689 Uttam Kumaran: it’s what’s all. It’s the fundamental technology is really, really magic. It is really this like orchestration and data problem. That is what a lot of firms are are now like, okay, how do we get around? We have, like all these sources, and even we have a lot of companies are like, I didn’t. We never even wrote down like what this process is. It sort of just happens like, or it’s like amorphous. Or that’s where we’re like, okay, we need like 3 meetings to just like.
211 00:34:37.250 ⇒ 00:34:41.279 Uttam Kumaran: do the question answer. And they’re like, Yeah, I guess I never thought about what I would expect.
212 00:34:41.420 ⇒ 00:34:50.260 Uttam Kumaran: But you have to treat it like an intern like if you don’t tell the intern what you’re expecting, right, they’re screwed right. And so that’s how we describe it to a lot of folks.
213 00:34:51.310 ⇒ 00:34:51.940 Anton Romash: Yeah.
214 00:34:51.949 ⇒ 00:34:52.999 Anthony: Makes sense.
215 00:34:53.600 ⇒ 00:34:54.230 Uttam Kumaran: Cool.
216 00:34:54.230 ⇒ 00:34:58.030 Anthony: Alright guys. Well, looking forward to exploring this, it’ll be a lot of fun.
217 00:34:58.430 ⇒ 00:35:00.770 Anton Romash: Yeah, thank you again for the time. Appreciate it.
218 00:35:00.770 ⇒ 00:35:02.700 Robert Tseng: Thank you. Anthony Anton.
219 00:35:02.860 ⇒ 00:35:04.140 Anthony: Thanks, guys, bye.
220 00:35:04.140 ⇒ 00:35:04.860 Uttam Kumaran: That’s it.