Meeting Title: Ad Tech AI Agents Consulting Date: 2025-09-26 Meeting participants: Javeed, Uttam Kumaran
WEBVTT
1 00:02:42.000 ⇒ 00:02:42.970 Uttam Kumaran: Hey, Javid.
2 00:02:44.000 ⇒ 00:02:45.200 Javeed: Hey, Ethan, how are you?
3 00:02:45.200 ⇒ 00:02:55.130 Uttam Kumaran: Hey, good, how are you? Give me one, sorry, I just… I’m just, traveling to a friend’s house, so… just getting on the Wi-Fi and everything, but, hope all is well.
4 00:02:55.600 ⇒ 00:02:57.000 Javeed: No, no worries.
5 00:02:58.550 ⇒ 00:03:00.879 Javeed: So you are based off of Austin, right?
6 00:03:00.880 ⇒ 00:03:06.850 Uttam Kumaran: I am based out of Austin. Yes, today I’m in Maryland, I’m in the East Coast, but I’m based out of Austin. I lived in… I live in East Austin.
7 00:03:07.130 ⇒ 00:03:08.440 Javeed: Oh, sounds good.
8 00:03:08.440 ⇒ 00:03:09.000 Uttam Kumaran: How about you?
9 00:03:09.000 ⇒ 00:03:09.600 Javeed: Okay.
10 00:03:10.050 ⇒ 00:03:11.150 Javeed: I’m in Dallas, actually.
11 00:03:11.340 ⇒ 00:03:15.350 Uttam Kumaran: Oh, okay, okay, great. Yeah, I may come to Dallas, actually,
12 00:03:15.710 ⇒ 00:03:19.490 Uttam Kumaran: Next week, or the week after, for a little, like,
13 00:03:19.700 ⇒ 00:03:24.569 Uttam Kumaran: there’s a Clay user group meeting, so in case I’m there, maybe I can say hi.
14 00:03:25.020 ⇒ 00:03:31.640 Javeed: Yeah, sure, just ping me, so we can have a coffee, or maybe lunch. You are coming in the daytime or evening?
15 00:03:31.640 ⇒ 00:03:38.510 Uttam Kumaran: It’ll be evening, so the thing is at 5 o’clock, but I’ll probably try to get there a little bit early, so maybe I’ll send you a note, yeah.
16 00:03:38.810 ⇒ 00:03:52.000 Javeed: That’ll work. Okay, so I know you have a short time, so I will be quick. Sure. Actually, we are working with a customer in an ad tech domain, so they have the LMS, Learning Management System.
17 00:03:52.000 ⇒ 00:04:00.289 Javeed: And, we are building a couple of agents, student agent and teacher agent, right? So, the purpose of these agents is
18 00:04:00.370 ⇒ 00:04:09.090 Javeed: to help navigate students where they are stuck. For example, if a student comes and asks, okay, what should I work today?
19 00:04:09.310 ⇒ 00:04:20.120 Javeed: Now, agents should go smartly, fetch all that data that is available, relevant data in the LMS, and take decisions accordingly, and recommend.
20 00:04:20.230 ⇒ 00:04:24.130 Javeed: So… We’ve built, like, kind of a…
21 00:04:24.590 ⇒ 00:04:33.999 Javeed: I would say a moderate-level agent that is somewhat working, but we tried multiple ways, like single agent, or…
22 00:04:34.000 ⇒ 00:04:35.199 Uttam Kumaran: Yeah, multi-agent system.
23 00:04:35.200 ⇒ 00:04:43.690 Javeed: Multi-agent, orchestration, I think now team is working on a customized workflow as well, because there’s a use cases, for example, if users
24 00:04:43.730 ⇒ 00:04:54.750 Javeed: select a project, the context should completely switch to a project. If it’s… they are asking about the, like, a math, I’ll switch. So yeah, I should remember, actually, based on the context, what I need to.
25 00:04:54.750 ⇒ 00:05:04.649 Javeed: help out, and there’s a lot of data in the LMS that need to be fed into the AI, so that’s where we are somewhat not able to meet the client expectations.
26 00:05:04.650 ⇒ 00:05:10.590 Uttam Kumaran: Okay, okay. And then on what KPIs are you guys missing on? Is it response time? Is it, like, accuracy?
27 00:05:11.700 ⇒ 00:05:16.190 Javeed: Accuracy. More of a, like, going into… into details about…
28 00:05:16.900 ⇒ 00:05:19.830 Javeed: What actually help a student required.
29 00:05:19.830 ⇒ 00:05:20.520 Uttam Kumaran: Okay, okay.
30 00:05:20.520 ⇒ 00:05:37.679 Javeed: So there’s a… one problem was the toning of the message, like, it should respond in a, like, a nicer way, so we are able to fine-tune, I think, 4.1 GPT, but the problem was 4.1 was not making efficient tools calling, because there’s.
31 00:05:37.680 ⇒ 00:05:38.610 Uttam Kumaran: Oh, yeah.
32 00:05:38.610 ⇒ 00:05:43.660 Javeed: tools available, so we have to use, like, now two models, like, we are using four for…
33 00:05:43.820 ⇒ 00:05:59.859 Javeed: the overall, toll calling, and then 4.1 for the constructing the message. So these type of techniques, we tried, but I feel like we are, might be not that much experience in it, so where we are struggling to pitch, to applic the right
34 00:06:00.120 ⇒ 00:06:01.160 Javeed: approach.
35 00:06:01.410 ⇒ 00:06:01.840 Uttam Kumaran: Okay.
36 00:06:01.840 ⇒ 00:06:14.989 Javeed: So is there anything you guys can help, either on the consulting part, or guiding the team, or might be some few hours a week, where you can go and see if the team is actually having the right implementation, or guiding them accordingly?
37 00:06:15.280 ⇒ 00:06:23.459 Uttam Kumaran: Yeah, definitely. The problem you described is something we do all the time. So we have a client, that’s in the home services space, where they have
38 00:06:23.700 ⇒ 00:06:38.880 Uttam Kumaran: tons of documents, both structured, unstructured. We not… but we not only help them, like, sort of structure some of that data, but we also build several tools. So, we have a tool that goes and calls a database, tool that pulls from, like, a 50-100 page document.
39 00:06:38.880 ⇒ 00:06:45.740 Uttam Kumaran: Tools that are pulling from spreadsheets. But it’s also the… the big thing that we did is we built, like, a golden data set before.
40 00:06:45.780 ⇒ 00:06:52.830 Uttam Kumaran: Right, we understood, like, what are the… what are their expectations for question and answers, and then we almost… we build, like, a little bit of a scoring system.
41 00:06:53.100 ⇒ 00:07:06.340 Uttam Kumaran: So at any point, you know, like, how far off you are, and then your team can continue to optimize. So certainly an area we can help. You know, I’m interested in sort of what the agent architecture was,
42 00:07:06.500 ⇒ 00:07:09.430 Uttam Kumaran: And, you know, we do a lot of stuff also on, like.
43 00:07:09.530 ⇒ 00:07:29.240 Uttam Kumaran: I feel like a lot of the tone and things like that, you just have to sort of get the prompt chaining right. Like, you can solve a lot of those. I don’t know necessarily whether you’ll need fine-tuning to do that, but you will have to guarantee that it works. So there are some tone-related things that we can totally share on how we orchestrated those prompts. Are you guys using any, like.
44 00:07:29.520 ⇒ 00:07:35.200 Uttam Kumaran: framework, is it all, like, in code, or are you using, like, N8N or another thing to build things?
45 00:07:35.200 ⇒ 00:07:36.710 Javeed: We are using LanChain, actually.
46 00:07:36.710 ⇒ 00:07:38.729 Uttam Kumaran: Okay, okay, okay, great, cool.
47 00:07:38.730 ⇒ 00:07:40.139 Javeed: and land better off, I think, yeah.
48 00:07:40.140 ⇒ 00:07:45.260 Uttam Kumaran: the UI… is it in a… it’s just in a UI that you guys built, or where does the output get to?
49 00:07:45.260 ⇒ 00:08:00.160 Javeed: Yeah, UI, we build it. The UI is, it’s like a… like a chatbot. You just ask a question and get the response. But the second use case is, which is still pending, is, like, the Najis use case. Like, AI should, like, identify
50 00:08:00.160 ⇒ 00:08:13.999 Javeed: Yes. What you need to tell to a student, like, they are calling it an adjust. For example, if a student is stuck somewhere in a math, right? So AI should send that reminder to teacher, this student is stuck with math and need help, right?
51 00:08:14.210 ⇒ 00:08:14.570 Uttam Kumaran: Yeah.
52 00:08:14.570 ⇒ 00:08:16.939 Javeed: These type of expectations are there.
53 00:08:16.940 ⇒ 00:08:24.690 Uttam Kumaran: And are you guys storing messages at all, or it’s just sort of single? Yes. Okay, okay, so then if you have the messages, then it’s totally possible to then pass all that through back to context.
54 00:08:24.760 ⇒ 00:08:36.880 Uttam Kumaran: to have. So we… for some of our systems, we actually have, sort of, two agents. We have something that is user-facing, and typically we have, like, a trainer, or in your case, like, a teacher, and there’s actually, like, a separate agent that serves them, because, of course.
55 00:08:36.880 ⇒ 00:08:45.139 Uttam Kumaran: the way they need to get outputs, and so definitely something that we can help with. I mean, I think kind of the way you described it is perfect. We have
56 00:08:45.430 ⇒ 00:08:50.950 Uttam Kumaran: some senior folks and some more mid-level folks, too. I’m happy to just, you know, if…
57 00:08:51.080 ⇒ 00:09:08.680 Uttam Kumaran: I’m happy to just… either one of them, I think, would be effective here. And if you have a sort of someone on your side, or however you want to arrange it, I think that’d be great. We have tons of examples that we can share. Like, we can sign an NDA and do all that, and then we can share exactly, like, how we’ve been doing it, for sure.
58 00:09:09.890 ⇒ 00:09:23.709 Javeed: Yeah, sure. So, definitely, the NDA is a must. Next up, you can send me the NDA and some kind of engagement model. I think you got the requirements. So, like, we have the team, like, that can help, that can do.
59 00:09:23.710 ⇒ 00:09:28.509 Uttam Kumaran: A little bit of technical consulting on, like, reviewing and providing, okay, like, what to try next.
60 00:09:28.780 ⇒ 00:09:29.330 Javeed: Yes.
61 00:09:29.330 ⇒ 00:09:33.699 Uttam Kumaran: What’s your timeline on it so far? Is it just, like, we need to get it now, basically?
62 00:09:34.070 ⇒ 00:09:35.780 Javeed: So, actually, we are already late.
63 00:09:35.780 ⇒ 00:09:36.970 Uttam Kumaran: So, okay.
64 00:09:37.130 ⇒ 00:09:47.690 Javeed: We spent 2 months, it’s around 8 weeks now, project. Initially, they were happy with the responses, and things were, like, coming into the right direction, but now they started getting frustrated.
65 00:09:47.690 ⇒ 00:09:51.179 Uttam Kumaran: And there’s no problems in the response times at all?
66 00:09:52.500 ⇒ 00:10:08.620 Javeed: But now their focus is, like, getting the right response. Even right now, we are not focusing on the timing part. We can definitely go and improve that, but first we need to, like, give them some confidence that AI started,
67 00:10:08.930 ⇒ 00:10:15.359 Javeed: giving the response that you want. So he… he’s… actually, he’s also too generic, but he generally say, okay, so AI should
68 00:10:16.210 ⇒ 00:10:35.999 Javeed: actually guide a student where a student is stuck, right? So how that can be achieved, what could be the questioning, answering, so that’s where we are also trying to figure it out. Another approach team suggested, I’m not sure if it’s the right way or not, to build actually a knowledge graph, like, using something like New 4J, and feed all that data, and then.
69 00:10:36.000 ⇒ 00:10:45.260 Uttam Kumaran: Yeah, I guess that was my… that was my question, is, like, for the LMS, like, is it just an AP… you just have API calls to certain docs, or, like, what is the structure of the underlying…
70 00:10:45.500 ⇒ 00:11:05.190 Javeed: So there’s a two type of data. We are using API, like MCP. We are creating tools, making API calls, and feeding that data. So there’s a two type of data. One is, like, a relational database, for example, the progress, how much the progress a student made. This is generally a student actually assigned projects. Projects have assessments, and the assessments actually have the scorings, right?
71 00:11:05.190 ⇒ 00:11:11.339 Javeed: Okay, yeah. So, the project is, like, more of an unstructured data, like, it includes the HTML, the content.
72 00:11:11.340 ⇒ 00:11:11.740 Uttam Kumaran: Okay.
73 00:11:11.740 ⇒ 00:11:25.359 Javeed: It includes links as well. So now, the project part, we are creating a vector embeddings. For every project, we have the vectors, but the relational data is coming through the API. Like, for example, if you want to see the progress.
74 00:11:25.580 ⇒ 00:11:27.279 Javeed: It’s a real-time data.
75 00:11:27.560 ⇒ 00:11:28.510 Uttam Kumaran: Okay, okay.
76 00:11:30.160 ⇒ 00:11:32.640 Uttam Kumaran: Okay, that makes sense. I think overall.
77 00:11:32.760 ⇒ 00:11:41.670 Uttam Kumaran: I sort of get the idea. I think that probably the… yeah, the best next step is I can try to get a contract over today, so that we can share some details. I think maybe I would love to introduce you
78 00:11:41.830 ⇒ 00:11:57.199 Uttam Kumaran: maybe to one… like, our team, and then we have someone… we have Sam on my team, he sort of leads a lot of our… our AI architecture, but also for this, if it’s… it seems like you guys are pretty far enough that, like, I may just pair you guys with one of my…
79 00:11:57.430 ⇒ 00:12:01.170 Uttam Kumaran: mid-level folks, that’s, like, we’re… we build a lot of these agents, they’re, like.
80 00:12:01.270 ⇒ 00:12:11.710 Uttam Kumaran: That’s what they do every day. So I think quickly, hopefully, in, like, one or two sessions, they can give you, like, try these, like, things, or we’ll share, like, on our end, things that are working.
81 00:12:13.070 ⇒ 00:12:19.100 Uttam Kumaran: And yeah, as you mentioned, just, like, a few hours a week, and if you guys are on Slack, then that’s probably easiest for us.
82 00:12:19.610 ⇒ 00:12:34.860 Javeed: Yes, so we are on Slack, I think that that’s fine. Like, we can connect the team. Mainly, from our side, it will be a one person. He can handle… the lead, actually, who is on the lead on this project. So it’s a three-member team right now, but the lead, I think he can take care, he can talk, and he can…
83 00:12:34.860 ⇒ 00:12:41.019 Javeed: Even he can show the actual architecture they have built right now, or what they previously tried, which worked, which didn’t work.
84 00:12:41.120 ⇒ 00:12:42.860 Uttam Kumaran: Great, okay, great.
85 00:12:43.010 ⇒ 00:12:52.989 Uttam Kumaran: Yeah, we’ve, we’ve dealt with a lot of these same issues, and I think we’ve just, like, tried a lot of different architectures, and I do a lot of reading in the market about, like, what’s working.
86 00:12:53.070 ⇒ 00:13:06.690 Uttam Kumaran: And so, a lot of our use cases are very knowledge-heavy. It’s not, like, simple chatbots. We have a ton of different tool calls, and, we do a lot of RAG-based use cases. But again, I think we found some shortcuts that are working.
87 00:13:06.690 ⇒ 00:13:13.219 Uttam Kumaran: And then also, like, I know the timeline’s quick, so you guys want to try to see some, like, really solid outputs, so you guys can keep going, so…
88 00:13:13.880 ⇒ 00:13:17.980 Javeed: Yeah, actually, they… even they asked, okay, share us something that… by Monday, that is working.
89 00:13:17.980 ⇒ 00:13:26.689 Uttam Kumaran: Okay. Alright, well, why don’t I… let me try to get something today, and I’ll tell you, sort of, like, what pricing is like just for a few hours, and then…
90 00:13:26.860 ⇒ 00:13:38.630 Uttam Kumaran: I mean, we’re… we can move as fast as you guys need, so, like, it… I mean, I don’t… even if they want to meet this weekend, that’s fine, if our guys are open to it, so whatever you guys need, just happy to help.
91 00:13:38.630 ⇒ 00:13:52.780 Javeed: Yeah, I think they’re working tomorrow, just to make the Monday deadline, but let’s share those, like, a contractual. So generally, just to give you an idea, so generally how that works, in terms of the price point and the commitment and all that.
92 00:13:53.070 ⇒ 00:13:53.550 Uttam Kumaran: Yes.
93 00:13:54.120 ⇒ 00:14:05.369 Uttam Kumaran: So, I mean, right now, I think for hourly, our range is around $200 an hour. This would be for, you know, just a normal engineer on our side, on the AI side.
94 00:14:05.410 ⇒ 00:14:17.400 Uttam Kumaran: I mean, like, for most of our typical clients, we like to assign larger engagements, like, month to month. This one is definitely more of, like, a strategy consulting, so I don’t mind doing hourly. Also.
95 00:14:17.400 ⇒ 00:14:30.529 Uttam Kumaran: like, I know you, so I don’t… it’s… I just want to be helpful. So, if you’re okay with that, I’m happy to just do a few hours, and then… certainly, actually, we should, you know, I think it’d be really nice to work together on this, and we should try to find more deals to collaborate on.
96 00:14:30.580 ⇒ 00:14:33.189 Uttam Kumaran: You know, so that would be…
97 00:14:33.190 ⇒ 00:14:38.469 Javeed: So there’s multiple deals in pipeline, but right now, I am not that confident we should take it.
98 00:14:38.470 ⇒ 00:14:39.000 Uttam Kumaran: Yeah.
99 00:14:39.000 ⇒ 00:14:44.010 Javeed: If we have something, have someone who can actually really deliver these type of…
100 00:14:44.240 ⇒ 00:14:51.680 Uttam Kumaran: Yeah, I think you’ll see what we’re doing from our side, you’ll kind of get a sense of that. And then, are you guys focused on education? Or is that, like, a clear area, or…
101 00:14:51.680 ⇒ 00:15:01.420 Javeed: No, so generally, our area is healthcare, but this is the customer we got. So, education and health, that we were previously working, but now we are purely focusing on health, actually.
102 00:15:01.420 ⇒ 00:15:02.520 Uttam Kumaran: Okay, okay.
103 00:15:02.520 ⇒ 00:15:13.509 Javeed: I have a good connection on the higher business leader side, so I generally get this type of questions, okay, so if we want to automate this business process through AI, how this will work?
104 00:15:13.510 ⇒ 00:15:24.599 Uttam Kumaran: Yeah, I mean, I think… I think next week, if I come to town, I’ll just show you, like, we are… we’re putting together a lot of materials, and that’s exactly what we’re working on for a lot of customers with, like, automating entry into ERP systems.
105 00:15:24.640 ⇒ 00:15:32.849 Uttam Kumaran: We build… a lot of our work is internal, so we’re building a ton of internal automations, and we’re selling directly to COO, CAO, so…
106 00:15:32.890 ⇒ 00:15:44.669 Uttam Kumaran: I think you… you may… we may find some opportunity as well, so I’ll share that, but I think first priority is, yeah, if… if around the $200 an hour price point seems okay, just for a few hours, you can… you can try to…
107 00:15:45.090 ⇒ 00:15:51.460 Uttam Kumaran: meet them… I can… I can try to get stuff over today, and if they want to meet tomorrow, even for an hour or two, that’s fine.
108 00:15:51.700 ⇒ 00:16:06.839 Javeed: Yeah, I think that’s fine. I think as a consulting strategy point of view, that will work. Although, like, our team is offshore, so for this project, our pricing point is very low, because it’s based on the offshore, but as a strategy, if we are able to, like, solve these things, that’s.
109 00:16:06.840 ⇒ 00:16:24.459 Uttam Kumaran: Yeah, I just want to make sure, and then also, like, we’re… we’re not, like, so strict, so, like, if we… if you have to message in Slack and things, like, we’re both consultants, so I’m not here to, like, you know… so I just want to make sure that you guys are supported, and then, yeah, hopefully, I think, hopefully we can help, and then find some more ways to make some money, so that’d be great.
110 00:16:24.990 ⇒ 00:16:41.850 Javeed: Yeah, let’s start with this, and see if we got a good value out of it, and then definitely I can start getting those via specific machine learning AI, specific related projects. We did a lot of chatbots and simple agents, but this use case is, like, hitting us.
111 00:16:41.850 ⇒ 00:16:47.360 Uttam Kumaran: Yeah, we’re doing some… we’re doing… we’re now… we started there, and then now some of our stuff is getting pretty, pretty…
112 00:16:47.460 ⇒ 00:16:50.749 Uttam Kumaran: big, like, way more significant, so…
113 00:16:51.010 ⇒ 00:16:58.569 Uttam Kumaran: Yeah, okay, perfect. So let me just get my team to start to send some stuff over. I’ll just start a quick Slack channel, and then get everything organized.
114 00:16:58.990 ⇒ 00:17:00.280 Javeed: Yeah, that should work.
115 00:17:00.280 ⇒ 00:17:01.689 Uttam Kumaran: Okay, okay. Thanks, Ravid.
116 00:17:01.690 ⇒ 00:17:03.130 Javeed: Appreciate it. Okay, then.
117 00:17:03.130 ⇒ 00:17:04.200 Uttam Kumaran: Okay, bye.