Meeting Title: Brainforge x Vixie (Wixel) Slack Assistant Discussion Date: 2026-02-11 Meeting participants: Uttam Kumaran, Samprita H
WEBVTT
1 00:00:56.250 ⇒ 00:00:58.329 Samprita H: Hey, Tim, can you hear me?
2 00:00:58.330 ⇒ 00:00:59.130 Uttam Kumaran: Yes.
3 00:01:00.660 ⇒ 00:01:02.230 Samprita H: Hey, nice to meet you.
4 00:01:03.080 ⇒ 00:01:03.779 Uttam Kumaran: Hey, how are you?
5 00:01:03.780 ⇒ 00:01:07.470 Samprita H: I’m really sorry. I don’t know why I didn’t get an alert or…
6 00:01:07.890 ⇒ 00:01:15.689 Uttam Kumaran: No, it’s okay, that’s fine, I just, like, this is also, like, I’m… I’m asking you a bunch of questions, so I… that’s no problem at all.
7 00:01:16.730 ⇒ 00:01:20.699 Samprita H: Okay, really, I really apologize.
8 00:01:21.100 ⇒ 00:01:22.640 Samprita H: No, no, no, no problem.
9 00:01:23.440 ⇒ 00:01:29.170 Uttam Kumaran: Yeah, I mean, we’re… we’re starting to build, you know, a kind of, like, a similar Slack assistant
10 00:01:29.390 ⇒ 00:01:32.940 Uttam Kumaran: To what y’all have done. I think we…
11 00:01:33.400 ⇒ 00:01:36.870 Uttam Kumaran: My… our angle is partly, like, having it
12 00:01:37.320 ⇒ 00:01:47.079 Uttam Kumaran: be able to recall things from, like, a pretty broad knowledge base, as well as some of the more, like, proactive things that, you know, Vixi does.
13 00:01:47.180 ⇒ 00:02:01.809 Uttam Kumaran: I’m just kind of interested in, like, how can we avoid some of the, like, dead ends that maybe you went through when you were sort of building it, and just, like, kind of curious a little bit about the architecture and, like, what you guys have learned so far.
14 00:02:03.770 ⇒ 00:02:20.220 Samprita H: Yeah, I mean, we have Wixie for, like, one, more than a year now, right? The thing is, most challenging thing that I, faced was keeping up with, the changes, offering in, in agentic frameworks, and…
15 00:02:20.660 ⇒ 00:02:22.160 Samprita H: Yeah. AI in general.
16 00:02:22.320 ⇒ 00:02:26.549 Samprita H: It changes so rapidly. By the time we roll out something.
17 00:02:26.750 ⇒ 00:02:30.750 Samprita H: AI world is totally different from what it was.
18 00:02:31.390 ⇒ 00:02:36.050 Uttam Kumaran: Yeah, exactly. No, I agree, even from, like, 2 years ago. Like, even a year ago, so…
19 00:02:36.050 ⇒ 00:02:36.720 Samprita H: Yeah.
20 00:02:37.080 ⇒ 00:02:37.480 Uttam Kumaran: Yeah.
21 00:02:37.480 ⇒ 00:02:45.599 Samprita H: So, we started with Langchain, and then came the agentic orchestration and all that. We switched, To…
22 00:02:46.290 ⇒ 00:02:56.810 Samprita H: We use OpenAI agent framework. I’m trying to keep the code phase as small as possible, so that… because I’m the only one who’s maintaining it.
23 00:02:57.120 ⇒ 00:02:57.740 Uttam Kumaran: Okay.
24 00:02:58.590 ⇒ 00:03:14.549 Samprita H: So, yeah, that’s another challenge. The third challenge, I think, because we… Wixie is, like, what you can call is a collection of agents. There are a bunch of specialist agents, and then there’s a,
25 00:03:14.740 ⇒ 00:03:23.390 Samprita H: One agent which, you know, kind of hands off the conversation to the specialist agent based on what the topic is about.
26 00:03:23.390 ⇒ 00:03:24.020 Uttam Kumaran: Okay.
27 00:03:24.730 ⇒ 00:03:25.320 Uttam Kumaran: And then…
28 00:03:25.750 ⇒ 00:03:30.429 Uttam Kumaran: And then I’m kind of curious, you know, I feel like you’re at… the agent you guys build is sort of the first
29 00:03:30.900 ⇒ 00:03:44.090 Uttam Kumaran: really great example of, like, something that’s more ambient. Like, it’s responding to people. Like, I’m wondering, like, how that mechanism works, and, like, are you guys reviewing before it goes out? Is there some type of.
30 00:03:44.090 ⇒ 00:03:45.780 Samprita H: Yeah, yeah, we do. Okay. Yeah.
31 00:03:46.030 ⇒ 00:03:48.689 Uttam Kumaran: Okay. So, we have.
32 00:03:48.790 ⇒ 00:04:00.670 Samprita H: every conversation in Slack, gets recorded, and we track, kind of, it analyzes the sentiment, whether… is this something, is that a request, or, something that needs to be responded?
33 00:04:00.670 ⇒ 00:04:01.940 Uttam Kumaran: Okay. 2?
34 00:04:01.950 ⇒ 00:04:16.879 Samprita H: If it’s, addressed to someone particular, for example, if it’s addressed to Vishnu, or Arthur, or Ali, you know, whoever, Vixi also keeps track of that, and it sends reminders if they haven’t responded in a certain.
35 00:04:16.880 ⇒ 00:04:18.349 Uttam Kumaran: Oh, really? Okay, and it DMs them.
36 00:04:18.350 ⇒ 00:04:19.679 Samprita H: them. Yeah.
37 00:04:19.870 ⇒ 00:04:20.470 Uttam Kumaran: Okay.
38 00:04:21.220 ⇒ 00:04:26.220 Samprita H: And if it can respond on its own, like, if it’s, if it’s a…
39 00:04:26.540 ⇒ 00:04:37.059 Samprita H: well, you could say, questions that already have answers in Wixel’s knowledge phase. Then it just formulates the answer, posts it internally first.
40 00:04:37.130 ⇒ 00:04:43.630 Uttam Kumaran: And, and you’re just using some simple retrieval to, like, find the relevant documents? Okay.
41 00:04:43.630 ⇒ 00:04:44.250 Samprita H: Yes.
42 00:04:44.480 ⇒ 00:04:48.069 Samprita H: We use, rag retrieval.
43 00:04:48.330 ⇒ 00:04:50.519 Samprita H: To get the answers.
44 00:04:51.050 ⇒ 00:05:00.640 Samprita H: And then it posts it. Someone can click the button, approve, or edit, or reject.
45 00:05:00.950 ⇒ 00:05:04.590 Samprita H: And… If someone approves it, then it gets posted.
46 00:05:04.590 ⇒ 00:05:10.770 Uttam Kumaran: Can… would you mind, like, showing me that flow? I’m just, like, interested in the UX within Slack.
47 00:05:10.960 ⇒ 00:05:13.219 Uttam Kumaran: If not, that’s also totally fine.
48 00:05:14.090 ⇒ 00:05:17.829 Samprita H: No, no, I can show you. Is it okay if I can share the screen?
49 00:05:17.830 ⇒ 00:05:19.129 Uttam Kumaran: Yeah, yeah, yeah, please.
50 00:05:19.660 ⇒ 00:05:38.199 Uttam Kumaran: So, I mean, because our, our, like, immediate roadmap is exactly that. It’s funny, because I’ve been wanting to build this for all, basically, like, a year, finally getting, like, bandwidth to work on it. But we’re gonna kind of do the same thing. We’re gonna have, like… we’re gonna start with, like, an approval flow process.
51 00:05:38.460 ⇒ 00:05:41.610 Uttam Kumaran: To, like, a group of us that can just approve. Okay.
52 00:05:41.750 ⇒ 00:05:45.149 Uttam Kumaran: We’re similarly, I’m gonna build sort of this reminder system.
53 00:05:46.050 ⇒ 00:05:49.850 Uttam Kumaran: And I’m gonna… I’m sorta gonna build it for me, and then roll it out.
54 00:05:50.230 ⇒ 00:05:51.829 Uttam Kumaran: Broadly to folks.
55 00:05:52.220 ⇒ 00:05:55.279 Samprita H: Our problem is, like, our knowledge base is, like.
56 00:05:55.930 ⇒ 00:05:57.240 Uttam Kumaran: really big.
57 00:05:57.400 ⇒ 00:06:05.769 Uttam Kumaran: like… we… You can think about it, we have almost, like, we record, like, every meeting, and so…
58 00:06:05.770 ⇒ 00:06:06.450 Samprita H: Okay.
59 00:06:06.450 ⇒ 00:06:11.969 Uttam Kumaran: It’s almost… it’s all those transcripts, it’s client documents, it’s web research, so…
60 00:06:12.170 ⇒ 00:06:21.900 Uttam Kumaran: sort of having to think a little bit about, like, how to formulate the answer, but I think we’ll figure that out. And then, yeah, I mean, the reminders thing, and then being able to
61 00:06:22.220 ⇒ 00:06:27.919 Uttam Kumaran: Proactively, sort of, give information is, like, is the next, like, two things we’re building.
62 00:06:29.050 ⇒ 00:06:29.720 Samprita H: Okay.
63 00:06:30.120 ⇒ 00:06:42.900 Samprita H: That’s really good. I mean, we are also facing the same problem, actually, with broader knowledge base. I mean, and knowledge sources are, like, from…
64 00:06:43.050 ⇒ 00:06:51.439 Samprita H: Different. There is… there are transcripts, there are one-on-one meetings, there are a lot of things, where the context is held.
65 00:06:53.210 ⇒ 00:06:58.789 Samprita H: And, you have to gather everything, before answering.
66 00:06:58.960 ⇒ 00:07:05.530 Uttam Kumaran: Yeah, the lovely thing is, like, we’ve done that already, like, we have it all in a database.
67 00:07:06.160 ⇒ 00:07:06.890 Samprita H: Yeah, okay.
68 00:07:06.890 ⇒ 00:07:12.520 Uttam Kumaran: vectorize. So, like, this is, like, Slack’s… blocks, transcripts.
69 00:07:12.730 ⇒ 00:07:19.150 Uttam Kumaran: And, like, we just recently are basically moving our whole company kind of into GitHub.
70 00:07:19.500 ⇒ 00:07:25.100 Uttam Kumaran: So, I’m kind of removing a lot of things out of Notion, and…
71 00:07:25.610 ⇒ 00:07:29.710 Uttam Kumaran: my hope is that, like, Notion is used more for, like.
72 00:07:30.120 ⇒ 00:07:35.639 Uttam Kumaran: policy and, like, HR and some notes, but, like, everything is gonna end up back in a GitHub repo.
73 00:07:36.220 ⇒ 00:07:39.200 Samprita H: Because everybody in the company is using Cursor now.
74 00:07:39.610 ⇒ 00:07:45.249 Uttam Kumaran: And using Cursor not only for engineering, but for, like, a lot of knowledge work, like sales…
75 00:07:45.250 ⇒ 00:07:45.839 Samprita H: Oh, okay.
76 00:07:45.840 ⇒ 00:07:53.050 Uttam Kumaran: marketing, And so, having, like, a single repo structure with all the information in there.
77 00:07:53.490 ⇒ 00:07:56.179 Uttam Kumaran: Is, like, been super, super helpful.
78 00:07:56.730 ⇒ 00:07:57.870 Uttam Kumaran: And so…
79 00:07:58.130 ⇒ 00:08:06.629 Uttam Kumaran: Yeah, we’re sort of, like, working now on, like, building up the knowledge for this agent. I mean, more of the issue in building Slack assistants is Slack, actually. It’s like…
80 00:08:07.330 ⇒ 00:08:11.840 Uttam Kumaran: It’s like, it’s so hard, like, one thing I’m like, how do we do end-to-end testing?
81 00:08:11.970 ⇒ 00:08:18.270 Uttam Kumaran: like, I want to be able to test locally, and then test, and then push to production, so like…
82 00:08:18.430 ⇒ 00:08:25.239 Uttam Kumaran: That’s what we’re figuring out, because it’s, like, sucks to, like, develop Slack assistants. Like, it’s hard to… it’s hard to do it.
83 00:08:25.940 ⇒ 00:08:38.900 Samprita H: In our case, you’re actually absolutely right. In our case, we have, like, two separate workspaces where we deploy to one workspace first, first, test it internally, and then, deploy.
84 00:08:38.900 ⇒ 00:08:40.340 Uttam Kumaran: Oh, okay.
85 00:08:40.460 ⇒ 00:08:51.589 Uttam Kumaran: Yeah, because I have a test assistant, I’m like, I can’t… well, I can’t, like, scope this to just a few people. It’s not working. Yeah. Because it’s installing the whole workspace, so… okay, that makes sense.
86 00:08:51.590 ⇒ 00:08:54.379 Samprita H: What you could do is, you could,
87 00:08:55.030 ⇒ 00:08:58.130 Samprita H: mimic the payload that gets sent to Slack.
88 00:08:59.290 ⇒ 00:08:59.940 Uttam Kumaran: Yeah.
89 00:09:00.500 ⇒ 00:09:03.460 Samprita H: And then, you know, tested from that.
90 00:09:04.530 ⇒ 00:09:06.839 Uttam Kumaran: Yeah, I wonder also… yeah, yeah.
91 00:09:07.600 ⇒ 00:09:13.140 Samprita H: I mean, capture the payload that gets, if you’re using, like, frameworks like Slack’s Bolt, or…
92 00:09:13.140 ⇒ 00:09:13.860 Uttam Kumaran: Yeah, we’re using Bulls.
93 00:09:13.860 ⇒ 00:09:21.299 Samprita H: standard framework. Okay. So, you just capture the exact payload that gets sent, and then,
94 00:09:21.700 ⇒ 00:09:24.209 Samprita H: You know, mimic that, smoked.
95 00:09:24.540 ⇒ 00:09:25.110 Uttam Kumaran: Okay.
96 00:09:28.550 ⇒ 00:09:33.820 Samprita H: Okay, let me, let me share my screen.
97 00:09:34.260 ⇒ 00:09:45.420 Samprita H: So, if you see Slack here, like, these are… these are… we have customer channels, right? We have, like, Slack responses that gets posted internally.
98 00:09:46.210 ⇒ 00:09:47.510 Samprita H: Do you see Ms. Green, by the way?
99 00:09:47.510 ⇒ 00:09:48.060 Uttam Kumaran: Yes.
100 00:09:48.500 ⇒ 00:09:53.600 Samprita H: Yeah, and, it gives us buttons, Whoever can approve, reject.
101 00:09:54.170 ⇒ 00:09:58.340 Samprita H: And then it gets posted to the channel where it was originally posted.
102 00:09:58.340 ⇒ 00:10:01.319 Uttam Kumaran: So if I look at… so there’s the original, and then there’s the…
103 00:10:02.440 ⇒ 00:10:05.730 Uttam Kumaran: Oh, the whole thing is the message, and when you, when you click edit.
104 00:10:05.830 ⇒ 00:10:07.939 Uttam Kumaran: It’s an inline editing, or how does it…
105 00:10:07.940 ⇒ 00:10:08.650 Samprita H: Yeah.
106 00:10:09.090 ⇒ 00:10:09.740 Uttam Kumaran: It pumps up.
107 00:10:09.740 ⇒ 00:10:13.479 Samprita H: inline, thingy.
108 00:10:15.820 ⇒ 00:10:17.430 Uttam Kumaran: Oh, awesome.
109 00:10:19.800 ⇒ 00:10:23.079 Samprita H: Or you can reach out. We also log all this.
110 00:10:23.490 ⇒ 00:10:27.330 Uttam Kumaran: Yeah, except for Jack, so you can kind of… yeah, I mean, my hope is that, like.
111 00:10:27.450 ⇒ 00:10:29.849 Uttam Kumaran: Over time, the stuff that’s, like.
112 00:10:30.140 ⇒ 00:10:35.010 Uttam Kumaran: 90% ac… because it’s all… for us, it’s internal, like, we’re not gonna put this on client stuff.
113 00:10:35.010 ⇒ 00:10:35.730 Samprita H: Okay.
114 00:10:37.540 ⇒ 00:10:38.459 Uttam Kumaran: So, I don’t be able to.
115 00:10:38.460 ⇒ 00:10:39.159 Samprita H: Keep track of everything.
116 00:10:39.160 ⇒ 00:10:40.689 Uttam Kumaran: More open. Okay.
117 00:10:40.690 ⇒ 00:10:44.520 Samprita H: back thread. We use Airtable to store our data.
118 00:10:44.820 ⇒ 00:10:49.270 Samprita H: So, we keep track of that Slack thread, at least the thread ID.
119 00:10:49.270 ⇒ 00:10:49.970 Uttam Kumaran: Great.
120 00:10:49.970 ⇒ 00:10:57.940 Samprita H: And then, what was the last conversation in that thread? Was that a request? Was… was that something that warranted a reply or not?
121 00:10:57.940 ⇒ 00:11:01.099 Uttam Kumaran: Yeah. And again, that is analyzed by AI, again.
122 00:11:01.450 ⇒ 00:11:01.900 Uttam Kumaran: Okay.
123 00:11:01.900 ⇒ 00:11:03.739 Samprita H: If it warrants a reply, then…
124 00:11:03.840 ⇒ 00:11:06.679 Samprita H: We get the relevant answer from our knowledge base.
125 00:11:07.150 ⇒ 00:11:08.290 Uttam Kumaran: I see, okay.
126 00:11:12.150 ⇒ 00:11:13.800 Uttam Kumaran: So, what’s on the roadmap?
127 00:11:14.750 ⇒ 00:11:17.430 Samprita H: To improve the responses,
128 00:11:17.430 ⇒ 00:11:17.810 Uttam Kumaran: Oh, okay.
129 00:11:17.810 ⇒ 00:11:19.370 Samprita H: Cope of the knowledge base.
130 00:11:19.370 ⇒ 00:11:19.910 Uttam Kumaran: Okay.
131 00:11:19.910 ⇒ 00:11:33.419 Samprita H: And because, you see, our goal is to make this, to answer… make it answer as… at least as well as humans, right? And humans store a lot of context.
132 00:11:33.520 ⇒ 00:11:43.900 Samprita H: when you’re talking to a person, some communication could have happened over email, something over Slack, something over a meeting as conversations.
133 00:11:43.900 ⇒ 00:11:53.790 Samprita H: So, how do you get all of that to AI? Right now, I think we have most of those… we have very limited context that’s available to AI.
134 00:11:53.880 ⇒ 00:11:55.470 Samprita H: So, expand the scope.
135 00:11:56.650 ⇒ 00:11:59.670 Uttam Kumaran: Yeah, if we can help with that, I mean, we’re gonna… we’re gonna…
136 00:11:59.970 ⇒ 00:12:03.420 Uttam Kumaran: get everything. So, like, I’m gonna pipe all of my emails in.
137 00:12:03.540 ⇒ 00:12:06.250 Uttam Kumaran: I’m gonna pipe all of our transcripts in.
138 00:12:06.670 ⇒ 00:12:10.630 Uttam Kumaran: and all of our, like, code changes as well.
139 00:12:12.920 ⇒ 00:12:19.779 Uttam Kumaran: even… even then, again, most of our problem that we’re dealing with is, like, communication with clients. It’s actually…
140 00:12:19.950 ⇒ 00:12:27.979 Uttam Kumaran: So most of the context, like, and this is, again, just kind of, like, our issue, is that, like, most of it is gonna be based on, like, reminders, and then, like.
141 00:12:28.080 ⇒ 00:12:31.959 Uttam Kumaran: It’s also a lot for me internally. For example, I want Vixie to be like, hey.
142 00:12:32.270 ⇒ 00:12:37.320 Uttam Kumaran: This, like, instead of asking for a meeting, this could be a loom.
143 00:12:37.940 ⇒ 00:12:46.959 Uttam Kumaran: Like, don’t… don’t schedule… don’t schedule this meeting you’re about to schedule. Like, that’s the sort of stuff that I am, like, trying to push. Like, I… I’m… like, it’s…
144 00:12:47.340 ⇒ 00:12:53.339 Uttam Kumaran: One thing that I do as a human is I’m everywhere, and I’m, like, giving people best practices, right? So, like, for example.
145 00:12:53.740 ⇒ 00:13:02.119 Uttam Kumaran: where is something? How did we do this before? Like, do we have any examples? Like, those are the types of questions.
146 00:13:02.680 ⇒ 00:13:11.529 Uttam Kumaran: That, like, I want this… this… we have, like, a brain forge assist, like, I want it to be able to answer, so I can push the best… so I can push…
147 00:13:11.710 ⇒ 00:13:16.500 Uttam Kumaran: the… the… the habits that we’re trying to build as a team, right? Like…
148 00:13:16.920 ⇒ 00:13:25.859 Uttam Kumaran: one of the big things is that, like, just, like, don’t book meetings. And so I often find myself, like, hey, this can just happen in Slack, or hey, record a quick loom.
149 00:13:26.040 ⇒ 00:13:37.360 Uttam Kumaran: Or for example, it’s like, hey, where do I find this thing? And again, even if we keep everything tight in Notion, whatever, people still are gonna ask, where is this, where is this? And so, being able to have an AI do that.
150 00:13:37.530 ⇒ 00:13:51.469 Uttam Kumaran: Our first thing we built was just like, hey, I’m gonna hook up Slack to 5.2, give it web search, and then now, anytime we’re having a discussion, I can, instead of going to ChatGPT, I can just inline the thread, ask a question.
151 00:13:52.620 ⇒ 00:13:56.489 Uttam Kumaran: So that was, like, our first start, but now we’re working on reminders.
152 00:13:56.740 ⇒ 00:13:59.080 Uttam Kumaran: And then we’re gonna work on something that’s more, like.
153 00:13:59.560 ⇒ 00:13:59.940 Samprita H: Yeah.
154 00:14:00.520 ⇒ 00:14:12.190 Samprita H: I mean, we have several agents. Internally, also, we do have agents, which help with marketing team, SEO optimization, we have agents, which help,
155 00:14:13.160 ⇒ 00:14:14.550 Samprita H: create,
156 00:14:14.720 ⇒ 00:14:24.139 Samprita H: You know, when cohort is running, when bootcamp training is running, to create the materials, to
157 00:14:24.250 ⇒ 00:14:28.450 Samprita H: Track the tasks that gets assigned to cohort participants.
158 00:14:28.450 ⇒ 00:14:29.090 Uttam Kumaran: Yeah.
159 00:14:29.370 ⇒ 00:14:35.249 Samprita H: To nudge them, to make sure that they complete it in time, or to make sure they don’t face any challenges.
160 00:14:35.450 ⇒ 00:14:43.660 Samprita H: I’ll soon… I think we have, like, or, like… 17, 18 agents so far.
161 00:14:43.820 ⇒ 00:14:44.400 Uttam Kumaran: Cool.
162 00:14:45.800 ⇒ 00:14:57.520 Samprita H: And, yeah, my other goal is to make it very easy to build new agents and, you know, integrate it with Wixie. Yeah. One of the challenges we face is,
163 00:14:58.230 ⇒ 00:15:11.969 Samprita H: coordinating with different teams, business teams, getting the prompts right, because they are the experts, business experts in those teams, right? So, getting that right is one other challenge.
164 00:15:12.230 ⇒ 00:15:20.370 Samprita H: How do I… for me as a developer, how do I make it easier for them to create their own agents?
165 00:15:21.630 ⇒ 00:15:22.130 Uttam Kumaran: Yeah.
166 00:15:22.130 ⇒ 00:15:26.799 Samprita H: and add it to Wixie, right? Just a few configuration files, or without having to touch code.
167 00:15:27.180 ⇒ 00:15:29.929 Samprita H: That is also something I’ve been trying to work on.
168 00:15:29.930 ⇒ 00:15:32.290 Uttam Kumaran: Yeah, like, I think for us, like.
169 00:15:33.960 ⇒ 00:15:36.240 Uttam Kumaran: Well, the way we’re thinking about it is, like.
170 00:15:36.640 ⇒ 00:15:43.109 Uttam Kumaran: I want humans in the right places, but most of our work actually is gonna end up being on, like, planning.
171 00:15:43.250 ⇒ 00:15:52.310 Uttam Kumaran: and requirements gathering, right? Because event… because basically what I’m trying to force our team to do is create a really clean environment to use tools like Codex.
172 00:15:52.480 ⇒ 00:15:53.990 Uttam Kumaran: Where we can kick off
173 00:15:54.670 ⇒ 00:16:05.210 Uttam Kumaran: development agent directly in Slack, but again, like, it’s… for… there’s two things that have to happen. One is, like, your requirements for codecs have to be really good.
174 00:16:05.340 ⇒ 00:16:07.630 Uttam Kumaran: The second piece is that Codex has to be able to
175 00:16:07.950 ⇒ 00:16:15.600 Uttam Kumaran: have end-to-end, like, development, right? You need to be able to run the app, test, run all the tests, push the PR…
176 00:16:15.600 ⇒ 00:16:17.280 Samprita H: Have that environment set up.
177 00:16:17.280 ⇒ 00:16:29.599 Uttam Kumaran: Yeah, and so that’s what we’re working… that’s what… that’s what we’re working on now. I think we are getting better and better, like, it’s… it’s not hitting as many age cases. And then the more of what I want my team to spend on is on planning.
178 00:16:29.720 ⇒ 00:16:47.499 Uttam Kumaran: So, like, I’m not gonna be able to get us… I don’t… I don’t actually want us to get out of, like, a business person, like, hey, I wish Fixie would do this. Like, okay, someone needs to take that, expand on that, get all the requirements, and then pass it to the AI to do the first pass, is, like, how we’re thinking about a lot of problems right now.
179 00:16:49.580 ⇒ 00:16:54.550 Samprita H: You could also think about having reusable epics or user stories.
180 00:16:54.610 ⇒ 00:16:55.150 Uttam Kumaran: Okay.
181 00:16:55.720 ⇒ 00:16:58.550 Samprita H: Get reused across multiple customers.
182 00:16:58.810 ⇒ 00:17:02.090 Samprita H: And have AI,
183 00:17:02.090 ⇒ 00:17:02.910 Uttam Kumaran: Do each of the time.
184 00:17:02.910 ⇒ 00:17:06.830 Samprita H: Right, use case, user story for you, based on what the requirement is.
185 00:17:07.010 ⇒ 00:17:22.159 Uttam Kumaran: Yeah, I guess I haven’t seen, like… some people are saying, like, hey, we’re gonna use AI to write the… to create the tickets, and then go one by one, but I’m sort of wondering, like, isn’t that what Codex is sort of doing under the hood? Like, it’s gonna break it up into tasks, and so…
186 00:17:22.790 ⇒ 00:17:25.770 Uttam Kumaran: I’m almost like… Yeah.
187 00:17:25.770 ⇒ 00:17:27.979 Samprita H: The high-level epics, or the high-level user stories.
188 00:17:28.319 ⇒ 00:17:32.829 Samprita H: That is visible to your customers. A customer might not be interested in the task picture.
189 00:17:32.830 ⇒ 00:17:36.829 Uttam Kumaran: So this is what I’m saying, is, like, this… none of this is gonna be customer-facing, for me.
190 00:17:36.830 ⇒ 00:17:37.380 Samprita H: Okay.
191 00:17:37.380 ⇒ 00:17:39.880 Uttam Kumaran: Meaning, like, this is all Brainforge internal.
192 00:17:40.640 ⇒ 00:17:46.099 Uttam Kumaran: Because the one thing is, as you know, like, we’re a consultancy, and so for us.
193 00:17:46.200 ⇒ 00:17:50.490 Uttam Kumaran: I’m not convinced yet that, like, our clients…
194 00:17:51.060 ⇒ 00:17:54.979 Uttam Kumaran: want the work done by AI, and want that to be known.
195 00:17:55.220 ⇒ 00:17:55.720 Samprita H: Oh, yeah.
196 00:17:55.720 ⇒ 00:17:59.720 Uttam Kumaran: I think they want us to move fast, and they want us to do a good job.
197 00:18:00.430 ⇒ 00:18:04.759 Uttam Kumaran: that means… that could be… mean a lot of things, right? So, partly that means…
198 00:18:04.920 ⇒ 00:18:10.009 Uttam Kumaran: I still need to… there still needs to be a smiling person behind the camera presenting work.
199 00:18:10.300 ⇒ 00:18:16.620 Uttam Kumaran: Partly we’re using AI, partly we’re still doing some manual work, but So far, the client, like…
200 00:18:17.740 ⇒ 00:18:20.829 Uttam Kumaran: it hasn’t been clear that the client’s like, oh, I wish, like.
201 00:18:20.960 ⇒ 00:18:25.769 Uttam Kumaran: we could just use your AI assistant. Like, they like working with us.
202 00:18:26.260 ⇒ 00:18:36.130 Uttam Kumaran: whatever we do in the background is up to us, right? And so this is why most of all of our agent work is gonna be behind a person.
203 00:18:36.590 ⇒ 00:18:39.180 Uttam Kumaran: Because, for us, like.
204 00:18:40.030 ⇒ 00:18:50.220 Uttam Kumaran: it’s not a differentiator that… it’s a… it’s more of a differentiator that we use AI to be fast than it is, like, you… our customer can use the AI. The customer… our customers.
205 00:18:50.220 ⇒ 00:19:00.390 Uttam Kumaran: are calling us because they don’t know something, right? They don’t… they have, like, no clue some of the things we’re building, and so part of our job… most of our job is the requirements gathering, is the coaching.
206 00:19:00.410 ⇒ 00:19:01.960 Uttam Kumaran: Is the communicating?
207 00:19:02.470 ⇒ 00:19:07.300 Uttam Kumaran: The execution piece, quite frankly, is actually more… the easy part of this job.
208 00:19:07.550 ⇒ 00:19:11.820 Uttam Kumaran: And that is what we’re offloading a lot to AI these days.
209 00:19:13.160 ⇒ 00:19:21.779 Uttam Kumaran: But for our… most of my use case is actually internal, which means it can be rough. Like, I don’t need to get it right, because I don’t need to get it perfect.
210 00:19:22.090 ⇒ 00:19:27.569 Uttam Kumaran: Because… it’s gonna be internal, right? So I can skip a lot of, like, steps.
211 00:19:27.860 ⇒ 00:19:33.389 Uttam Kumaran: Meaning, like, internally, we all have a really clear understanding of, like, what our business is.
212 00:19:33.560 ⇒ 00:19:38.030 Uttam Kumaran: And I don’t need it to be, like, Perfect to get it out.
213 00:19:38.860 ⇒ 00:19:39.300 Samprita H: Yeah.
214 00:19:39.300 ⇒ 00:19:44.599 Uttam Kumaran: Right, so I… somewhat I have that… I have that luxury going for us, in that, like, it could be a little bit rough.
215 00:19:45.050 ⇒ 00:19:48.669 Uttam Kumaran: Because, like, I’m not… I’m not worried about, like, someone internally being, like.
216 00:19:49.320 ⇒ 00:19:58.970 Uttam Kumaran: why did it miss the answer? It’s like, you know, I’m just… I’m working on it, so… versus if it’s client-facing, then it has… yeah, the level of rigor is much higher.
217 00:19:58.970 ⇒ 00:19:59.530 Samprita H: Yep.
218 00:19:59.950 ⇒ 00:20:00.550 Uttam Kumaran: Yeah.
219 00:20:01.850 ⇒ 00:20:06.730 Samprita H: And we need to have suitable gatekeepers or, you know, people who send it out.
220 00:20:06.730 ⇒ 00:20:21.029 Uttam Kumaran: Like, for example, I will be a lot more liberal with AI responding without intervention, quickly. Yeah. Like, I will… I’ll do the… I’ll do the approval flow for a bit, and then eventually I will…
221 00:20:21.900 ⇒ 00:20:26.009 Uttam Kumaran: Start to… start to pick off the top 20% and be like, just ship it.
222 00:20:26.380 ⇒ 00:20:42.810 Uttam Kumaran: Because something… some of the questions, I feel like it’s gonna be able to just answer, and it’s gonna be basics, and so I’m actually gonna be much more liberal with it, because it’s, again, it’s internal, right? So… Yeah. If AI gets it slightly wrong, I actually can see the feedback.
223 00:20:43.010 ⇒ 00:20:45.970 Uttam Kumaran: And the risk is… the risk is a lot lower, you know?
224 00:20:45.970 ⇒ 00:20:46.630 Samprita H: Yeah.
225 00:20:46.630 ⇒ 00:20:49.120 Uttam Kumaran: So that’s how we’re thinking about a lot more, like.
226 00:20:50.400 ⇒ 00:20:53.259 Uttam Kumaran: But I don’t know, it may change.
227 00:20:53.650 ⇒ 00:20:57.680 Uttam Kumaran: Yeah.
228 00:21:00.270 ⇒ 00:21:06.309 Samprita H: at least, at least if AI can point, the users to the right source, That is still helpful.
229 00:21:06.310 ⇒ 00:21:09.770 Uttam Kumaran: Yes. Or it’s like, hey, I don’t know, but it could be here.
230 00:21:10.070 ⇒ 00:21:13.390 Uttam Kumaran: Or, like, there’s information here, I’m not exactly sure.
231 00:21:13.720 ⇒ 00:21:14.550 Uttam Kumaran: Yeah.
232 00:21:15.590 ⇒ 00:21:21.569 Uttam Kumaran: So yeah, we, I mean, for when we do a lot of AI work, we start running a lot of, like, evals, basically.
233 00:21:21.960 ⇒ 00:21:24.520 Uttam Kumaran: So, we will… we’ll create, like.
234 00:21:25.300 ⇒ 00:21:28.319 Uttam Kumaran: for example, like, I have a laundry list of the…
235 00:21:28.420 ⇒ 00:21:32.409 Uttam Kumaran: easy, medium, like, hard questions that I want AI to answer.
236 00:21:32.580 ⇒ 00:21:35.620 Uttam Kumaran: Another thing that I’m gonna probably do is, like, have…
237 00:21:36.050 ⇒ 00:21:42.260 Uttam Kumaran: a step before this project, like, have AI go through and classify a lot of our Slack messages.
238 00:21:42.490 ⇒ 00:21:45.280 Uttam Kumaran: So we can start to build some type of cohort, like…
239 00:21:45.760 ⇒ 00:21:52.569 Uttam Kumaran: 80% of the messages are, like… or 30% are logistical, 40% are, like… Communication-based, third…
240 00:21:52.570 ⇒ 00:21:53.580 Samprita H: We actually do that.
241 00:21:53.580 ⇒ 00:21:54.120 Uttam Kumaran: Okay, cool.
242 00:21:54.120 ⇒ 00:21:55.150 Samprita H: have a workflow.
243 00:21:55.270 ⇒ 00:22:02.679 Samprita H: Like, if there’s a Slack conversation, that’s really useful. I mean, if we think we could… other…
244 00:22:03.300 ⇒ 00:22:03.690 Uttam Kumaran: Yeah.
245 00:22:04.080 ⇒ 00:22:15.259 Samprita H: we could get real insights from it. It could add to our knowledge base. We just, there’s a push button, we just push it, it gets sanitized, and
246 00:22:15.380 ⇒ 00:22:18.120 Samprita H: A knowledge base gets created from it.
247 00:22:18.380 ⇒ 00:22:18.930 Uttam Kumaran: Cool.
248 00:22:20.210 ⇒ 00:22:26.519 Samprita H: I mean, AI can be used for data generation as well, not just for consumption.
249 00:22:27.130 ⇒ 00:22:28.760 Uttam Kumaran: Mmm. Right? Yeah.
250 00:22:32.310 ⇒ 00:22:37.760 Samprita H: if you happen to see a question in Slack that gets, you want to just add it to FAQ.
251 00:22:38.910 ⇒ 00:22:39.380 Uttam Kumaran: Yeah.
252 00:22:39.380 ⇒ 00:22:43.979 Samprita H: workflows page, just push a button, AI creates an FAQ from that conversation and adds it.
253 00:22:44.410 ⇒ 00:22:45.130 Uttam Kumaran: Okay.
254 00:22:48.140 ⇒ 00:22:54.100 Samprita H: Because our initial challenge was to, how do we maintain the data? How do we maintain the knowledge base?
255 00:22:54.100 ⇒ 00:22:56.860 Uttam Kumaran: How… Yeah, you need to loop, yeah.
256 00:22:56.860 ⇒ 00:23:04.300 Samprita H: Yeah, yeah, and it required a lot of, effort, human effort, to create new guides, a new knowledge base.
257 00:23:04.810 ⇒ 00:23:13.759 Samprita H: If we made it easier, okay, you found something, for this use case, okay, this is very useful, this approach is useful, you want to document it somehow?
258 00:23:14.390 ⇒ 00:23:17.239 Samprita H: use AI to make it easy to document that.
259 00:23:18.180 ⇒ 00:23:18.980 Uttam Kumaran: Yeah.
260 00:23:19.610 ⇒ 00:23:24.240 Samprita H: That way, internal users,
261 00:23:25.250 ⇒ 00:23:28.239 Samprita H: Tend to document things more if it’s easier for them.
262 00:23:28.700 ⇒ 00:23:30.439 Samprita H: That’s what happened.
263 00:23:31.030 ⇒ 00:23:38.300 Uttam Kumaran: Yeah, so one thing could be, like, hey, I noticed that this conversation happened, there… it… there’s nothing related to this in the…
264 00:23:38.700 ⇒ 00:23:43.110 Uttam Kumaran: in the knowledge base, should I… should I go ahead and write that? Yeah.
265 00:23:45.830 ⇒ 00:23:47.810 Samprita H: Or should I have AI write it?
266 00:23:47.810 ⇒ 00:23:52.349 Uttam Kumaran: No, that’s, so that’s exactly it, like, because for us, like, that could be easily done, like.
267 00:23:52.510 ⇒ 00:23:56.890 Uttam Kumaran: go ahead and create a PR with the new Markdown file for this.
268 00:23:56.890 ⇒ 00:23:57.420 Samprita H: Yeah.
269 00:23:58.470 ⇒ 00:24:03.039 Uttam Kumaran: Of course, it’ll get reviewed by a human. Yeah, yeah, yeah. AI will do the initial legwork.
270 00:24:04.910 ⇒ 00:24:05.590 Uttam Kumaran: Yeah.
271 00:24:10.660 ⇒ 00:24:19.430 Samprita H: I think the other goal that I have is, I’m looking into eval sets, eval framework, like,
272 00:24:19.980 ⇒ 00:24:28.589 Samprita H: have a set of evaluation, like, these are the questions, these are the expected answers, and every time you want to change AI models or compare between two models.
273 00:24:28.770 ⇒ 00:24:30.100 Samprita H: You could just run it.
274 00:24:30.250 ⇒ 00:24:34.899 Uttam Kumaran: Yeah, we use, I think we’re using BrainTrust for a lot of evals.
275 00:24:35.470 ⇒ 00:24:35.940 Samprita H: Okay.
276 00:24:35.940 ⇒ 00:24:38.510 Uttam Kumaran: And then we also are using LankFuse.
277 00:24:38.960 ⇒ 00:24:40.710 Uttam Kumaran: Found those to be really good.
278 00:24:40.970 ⇒ 00:24:44.350 Uttam Kumaran: Yeah, for our clients, we’re doing a lot of eval work.
279 00:24:45.180 ⇒ 00:24:49.729 Uttam Kumaran: It just takes time, like, you have to build a really robust data set, and…
280 00:24:49.730 ⇒ 00:24:50.780 Samprita H: Yeah, yeah, that’s true.
281 00:24:50.780 ⇒ 00:24:57.990 Uttam Kumaran: It takes a lot, it takes… it just takes time, so… But, like, again, I think… Once we get…
282 00:24:57.990 ⇒ 00:25:00.529 Samprita H: one-time investment, right? Once you do it. Yeah, yeah.
283 00:25:00.530 ⇒ 00:25:03.420 Uttam Kumaran: Usually a one-time investment. Upgrade the models, and otherwise.
284 00:25:03.420 ⇒ 00:25:07.150 Samprita H: For us, changing the model takes a lot of time.
285 00:25:08.130 ⇒ 00:25:11.310 Uttam Kumaran: Yeah, no, it’s totally a one-time investment. It’s also, like, how do you…
286 00:25:11.480 ⇒ 00:25:15.120 Uttam Kumaran: If you need to make updates to system prompts and things, how do you test
287 00:25:15.300 ⇒ 00:25:17.339 Uttam Kumaran: That those are actually working.
288 00:25:20.640 ⇒ 00:25:30.309 Uttam Kumaran: Yeah, and so that’s what we’ll… like, once we kind of get, I think, the core functionality, and it starts to work, I’ll then start to layer on more evaluations.
289 00:25:30.450 ⇒ 00:25:33.869 Uttam Kumaran: So that we can cycle through, like, all of the models.
290 00:25:34.300 ⇒ 00:25:38.670 Uttam Kumaran: And, like, understand, like, cost impacts, like, do some type of routing.
291 00:25:38.910 ⇒ 00:25:42.280 Uttam Kumaran: Yeah. Yeah.
292 00:25:45.920 ⇒ 00:25:54.810 Samprita H: Because, as in, like, it is what we’re seeing, more and more agents you add, each agent’s, each agent might perform well with different
293 00:25:55.080 ⇒ 00:25:56.340 Samprita H: LLM model.
294 00:25:56.340 ⇒ 00:25:57.170 Uttam Kumaran: Yes.
295 00:25:57.170 ⇒ 00:26:01.159 Samprita H: Some might require a lot of reasoning, some doesn’t require it.
296 00:26:01.290 ⇒ 00:26:05.010 Samprita H: So… That’s… that’s my next goal, too.
297 00:26:05.330 ⇒ 00:26:09.820 Samprita H: Create a solid eval framework, and have a good set of evaluation.
298 00:26:10.740 ⇒ 00:26:15.110 Uttam Kumaran: Yeah, another thing is you can basically also, like, score the response.
299 00:26:15.160 ⇒ 00:26:19.630 Samprita H: Yeah. And when you send the… when you send the note in, so it gives you a sense of, like.
300 00:26:20.910 ⇒ 00:26:24.410 Uttam Kumaran: Do you feel like this response is good? Does it warrant, like, an edit?
301 00:26:24.730 ⇒ 00:26:26.570 Uttam Kumaran: But then also, again, like, I think
302 00:26:26.760 ⇒ 00:26:28.770 Uttam Kumaran: For us, when we work with clients, it’s…
303 00:26:28.990 ⇒ 00:26:35.860 Uttam Kumaran: the evals are something that we report on, so we want to see scores go up, and we want to see response times go down, right? So…
304 00:26:36.080 ⇒ 00:26:37.950 Uttam Kumaran: We report out on, like.
305 00:26:38.670 ⇒ 00:26:54.960 Uttam Kumaran: When we start, like, we don’t expect our scores to be good, and then as we grow and we get better, like, we are able to show that, like, okay, our average scores are going up, like, our accuracy is getting better, and then both those tools provide a lot of different, like, scoring frameworks.
306 00:26:54.960 ⇒ 00:26:55.640 Samprita H: Yeah.
307 00:27:00.490 ⇒ 00:27:09.000 Uttam Kumaran: Great. This was so helpful. Yeah, I feel like I have a lot of ideas, so I’ll… I’ll Slack you as we’ve kind of, like, shipped a couple things. You can let me know what you think.
308 00:27:09.550 ⇒ 00:27:15.620 Uttam Kumaran: Yeah, and then if I can be helpful, if you’re working on… if you need any help on anything, we do a lot of work with, like.
309 00:27:15.930 ⇒ 00:27:29.569 Uttam Kumaran: ton of, like, vector database. The problem with this project is that, like, it’s internal, so I’m working on it. All of our AI folks are working on client stuff, so I have to go borrow their time, and I’m like, hey, please, I’m stuck, help me out.
310 00:27:29.750 ⇒ 00:27:36.879 Uttam Kumaran: But, I don’t know, I’ve gone very far. Like, I know all the core concepts…
311 00:27:36.880 ⇒ 00:27:44.829 Samprita H: How… I have a question for you, like, since you work with Vector, how effective is vector database versus, just agentic tool calls?
312 00:27:45.270 ⇒ 00:27:46.339 Samprita H: Like… It’s right.
313 00:27:46.340 ⇒ 00:27:51.200 Uttam Kumaran: It kind of depends on, like, what you’re doing. Like, for example, for transcript search.
314 00:27:51.630 ⇒ 00:27:55.019 Uttam Kumaran: It’s probably best that you do some level of, like.
315 00:27:55.150 ⇒ 00:28:00.410 Uttam Kumaran: Keyword search to get stuff that’s relevant, and then pull that into context.
316 00:28:00.830 ⇒ 00:28:01.750 Samprita H: Okay.
317 00:28:01.920 ⇒ 00:28:03.979 Uttam Kumaran: In terms of tool calls, like.
318 00:28:04.100 ⇒ 00:28:23.150 Uttam Kumaran: I’m actually much more biased towards just using CLI where we… where possible. The nice thing about CLI is, like, the agent can literally write, like, CLI, like, for example, let’s say you’re using, like, the… we use Railway for some hosting. Let’s say you’re using Railway CLI. The agent can say CLI help and see everything it can do.
319 00:28:23.520 ⇒ 00:28:33.570 Uttam Kumaran: Versus MCP, there is a lot of, like, context stuffing that happens when you’re using MCP. Has to pull all those tools into context, all of the…
320 00:28:33.570 ⇒ 00:28:36.309 Samprita H: JSON descriptions of what they do. Oh, okay.
321 00:28:36.310 ⇒ 00:28:40.819 Uttam Kumaran: And so, it can tend to, like, Noise up the context layer.
322 00:28:40.820 ⇒ 00:28:41.450 Samprita H: Okay.
323 00:28:42.180 ⇒ 00:28:56.950 Uttam Kumaran: So instead, like, what we do is, where possible, we will rely on CLI, and then where not possible, like, Linear doesn’t have a… they have only a GraphQL API, so for… in that situation, we’ll use the MCP, and I’ll just be a little bit more mindful of the tools.
324 00:28:58.880 ⇒ 00:29:06.929 Uttam Kumaran: for, like, for coding agents, like, for codecs, for example, like, I’m gonna rely as much on CLIs as possible.
325 00:29:07.200 ⇒ 00:29:25.570 Uttam Kumaran: But here’s a good example. Like, HubSpot, right? We wrote… HubSpot has an MCP, but it’s only read-only. And so what I did is I’m like, okay, then I need to write back also. So then I wrote some helper functions, and I created some, like, a helper function… set of functions that allow people
326 00:29:25.750 ⇒ 00:29:36.839 Uttam Kumaran: when they’re saying, hey, look up this deal, update this field, it then switches to use my API functions to go run those. Like, modify deal, modify property, things like that.
327 00:29:37.300 ⇒ 00:29:46.069 Uttam Kumaran: It’s clunky, like, but you can tell HubSpot, the reason why is that if you… if they release a write MCP, I’m never gonna log into the UI ever again.
328 00:29:46.450 ⇒ 00:29:50.699 Uttam Kumaran: So that’s why these guys are playing, like, games slowly. It kind of sucks.
329 00:29:51.010 ⇒ 00:29:54.639 Uttam Kumaran: Because a lot of these tools are realizing that they are…
330 00:29:54.980 ⇒ 00:29:57.649 Uttam Kumaran: they’re just a UI on, like, a simple database.
331 00:29:57.650 ⇒ 00:29:58.190 Samprita H: Yeah.
332 00:29:58.190 ⇒ 00:30:03.369 Uttam Kumaran: And, I’m excited to stop paying for them as soon as possible.
333 00:30:03.660 ⇒ 00:30:10.510 Uttam Kumaran: like, I’m fairly… I’m really, like, a little bit, like, ruthless on this, and, like, I’m trying to find, like, as many tools
334 00:30:10.730 ⇒ 00:30:17.849 Uttam Kumaran: that can’t just be… that we’re using, where I’m like, is this a light abstraction on, like, a simple relational database? Like, why are we using…
335 00:30:18.330 ⇒ 00:30:18.820 Samprita H: Yeah.
336 00:30:18.820 ⇒ 00:30:22.560 Uttam Kumaran: Can I… can you… can I just get you to use… like, can I… for example.
337 00:30:22.820 ⇒ 00:30:32.110 Uttam Kumaran: I can… you could just generate, like, a UI on the fly, like a CRM database UI on the fly, right? So, why not, when you need that, just generate it, like.
338 00:30:32.280 ⇒ 00:30:34.789 Uttam Kumaran: And part of this is, like, I think education, like…
339 00:30:35.160 ⇒ 00:30:38.670 Uttam Kumaran: Not everybody in the company is trained on using Kursor.
340 00:30:39.170 ⇒ 00:30:46.700 Uttam Kumaran: and things like that. I’m forcing that. Like, I’m basically being like, there’s no alternative. Like, we are gonna consistently remove tools.
341 00:30:46.940 ⇒ 00:30:52.609 Uttam Kumaran: And all knowledge work is going to start to have to happen in tools like Cursor.
342 00:30:53.050 ⇒ 00:31:06.279 Uttam Kumaran: Because it’s so much faster for you to ask a question, hey, tell me about a time where we worked on this type of project for a client. Cursor will go and it has indexed thousands of files, transcripts, and will build you what you need.
343 00:31:06.380 ⇒ 00:31:07.540 Uttam Kumaran: When you need it.
344 00:31:07.650 ⇒ 00:31:10.479 Uttam Kumaran: And I almost want our company to, like, s…
345 00:31:10.650 ⇒ 00:31:14.469 Uttam Kumaran: move past ChatGPT and Claude for work quickly.
346 00:31:14.790 ⇒ 00:31:20.410 Uttam Kumaran: Like, those are nice, but it’s assuming that your… your team is, like, non-technical.
347 00:31:20.690 ⇒ 00:31:26.210 Uttam Kumaran: Right? And they can only work with, like, file… I’m like, no, everybody’s using Cursor. If you’re at the company, you’re using Cursor.
348 00:31:26.440 ⇒ 00:31:33.330 Uttam Kumaran: Like, most of our company’s engineers, and so they have no problem, and my job is actually for the business folks to get them in there.
349 00:31:33.810 ⇒ 00:31:39.449 Uttam Kumaran: And they’re doing it. And so, it’s, like, sort of a non-starter anymore. And so…
350 00:31:40.210 ⇒ 00:31:47.439 Uttam Kumaran: I think long story short, yeah, like, some things we’re still using… we’re using a mix of MCP, CLI or API for, but it’s…
351 00:31:47.550 ⇒ 00:31:54.129 Samprita H: it’s… it’s… it’s changing as fast as, like, I can… I can change. That’s the challenge, yeah, it’s changing, like…
352 00:31:56.100 ⇒ 00:32:02.489 Uttam Kumaran: by the morning you wake up, things… things are totally different than what you had the night before. Yeah, yeah. So that’s why more important to me is, like.
353 00:32:02.920 ⇒ 00:32:06.119 Uttam Kumaran: I wanna… my job at the company is to, like.
354 00:32:07.000 ⇒ 00:32:18.889 Uttam Kumaran: is to not only set the vision, but then, like, some of these things are so esoteric that a lot of people are not gonna get this conversation if I’m talking about, like, I’ve just thought about this for… so my job is to do the first version.
355 00:32:19.020 ⇒ 00:32:29.789 Uttam Kumaran: the moment someone asks, hey, that was great, I wonder if I can build this, I’m like, no, you can go do that. Here’s how to go do that. So, like, that’s my job, because I have no… my limitation is…
356 00:32:29.970 ⇒ 00:32:34.550 Uttam Kumaran: how many hours my eyes can stay open, you know? And so…
357 00:32:34.710 ⇒ 00:32:37.560 Uttam Kumaran: Which, which is, like, really tough, so…
358 00:32:37.660 ⇒ 00:32:54.319 Uttam Kumaran: for me, I have to almost come in and be… people aren’t… don’t understand when I’m like, Slack is gonna be able to do this what, so I’m like, I’m just gonna do it, you’re gonna see it. The moment there’s a feature request, okay, now you can go handle that, like, either via Codex, or you go ship the first version.
359 00:32:54.520 ⇒ 00:33:01.070 Uttam Kumaran: I will review, or the AI team will review your PR. And, like, that’s what I’m sort of pushing our organization towards.
360 00:33:01.450 ⇒ 00:33:06.280 Uttam Kumaran: But it’s… it’s challenging. It’s like a everyday, like, pushing, like…
361 00:33:06.280 ⇒ 00:33:10.669 Samprita H: No, it requires a whole mindset change, more than skills, I think.
362 00:33:10.670 ⇒ 00:33:11.210 Uttam Kumaran: Yeah.
363 00:33:13.080 ⇒ 00:33:16.079 Samprita H: AI-first thinking is a totally different mindset.
364 00:33:16.080 ⇒ 00:33:18.269 Uttam Kumaran: Yeah, and you have to be, you have to be pretty relentless.
365 00:33:18.640 ⇒ 00:33:24.479 Uttam Kumaran: Yeah, and it used to be very annoying. Like, I’m, like, I’m very, very annoying with people being like.
366 00:33:24.730 ⇒ 00:33:38.369 Uttam Kumaran: Did you go ask Cursor? Did you open Cursor and ask this question? Because people will be like, I wonder, like, how did we do this at the company? And I’m like, why are you asking this in Slack? Because… and then also, some people will go call engineers.
367 00:33:38.480 ⇒ 00:33:48.069 Uttam Kumaran: And that’s when I get very defensive, because I’m like, you are ripping an engineer who is working on a client out for something internally, causing a contact switch.
368 00:33:48.620 ⇒ 00:33:51.709 Uttam Kumaran: Like, how much… what was the… what was the cost of that?
369 00:33:52.120 ⇒ 00:33:54.940 Uttam Kumaran: That could have been, like, $500,000.
370 00:33:55.050 ⇒ 00:33:57.529 Uttam Kumaran: Context interruption that would just happen there.
371 00:33:57.840 ⇒ 00:34:02.200 Uttam Kumaran: And it’s actually the internal folks that are having the toughest, because the engineers
372 00:34:02.500 ⇒ 00:34:06.089 Uttam Kumaran: we adopted Cursor and all these things day one, right? It’s…
373 00:34:06.830 ⇒ 00:34:10.250 Uttam Kumaran: It’s the business folks that are having the most challenge
374 00:34:11.239 ⇒ 00:34:16.319 Uttam Kumaran: Because they never saw their work as, like, functions and, like, inputs and outputs.
375 00:34:16.909 ⇒ 00:34:20.140 Uttam Kumaran: And I have to for… I have to really push that, like, I’m like…
376 00:34:21.010 ⇒ 00:34:24.589 Uttam Kumaran: Anytime you have a question, you should… you should ask AI first.
377 00:34:25.449 ⇒ 00:34:29.629 Uttam Kumaran: or when you come to the table with a question, you should be like, here’s how I use AI to…
378 00:34:30.010 ⇒ 00:34:31.420 Uttam Kumaran: the answer, and then…
379 00:34:31.719 ⇒ 00:34:41.960 Uttam Kumaran: again, my annoyance is gonna get into the Slack, like, Slack will respond to you and be like, did you ask AI for this? Or it’ll just give you the answer. So that’s how I will have to solve it, because…
380 00:34:42.090 ⇒ 00:34:44.440 Uttam Kumaran: Getting kind of impatient.
381 00:34:46.389 ⇒ 00:34:47.260 Uttam Kumaran: Yeah.
382 00:34:49.610 ⇒ 00:34:57.259 Uttam Kumaran: Cool, okay. Well, I appreciate the time, yeah, thank you so much, and then, yeah, I’ll send you some notes as we sort of, like, are figuring this out as we go.
383 00:34:57.260 ⇒ 00:35:02.029 Samprita H: Yeah, yeah, you can Slack me anytime. I might ping you if I need some help with…
384 00:35:02.030 ⇒ 00:35:02.670 Uttam Kumaran: Perfect.
385 00:35:02.900 ⇒ 00:35:04.990 Samprita H: larger AI.
386 00:35:06.450 ⇒ 00:35:07.449 Uttam Kumaran: Yeah, no problem at all.
387 00:35:07.450 ⇒ 00:35:10.100 Samprita H: I’m still catching up, so…
388 00:35:10.100 ⇒ 00:35:14.560 Uttam Kumaran: Yeah, we are, like, I feel lucky we have 4 or 5 people that are…
389 00:35:14.780 ⇒ 00:35:19.010 Uttam Kumaran: full-time AI folks, but again, they’re all working… a lot of them are working on client stuff.
390 00:35:19.430 ⇒ 00:35:30.140 Uttam Kumaran: But the funny thing is our clients are not asking for stuff nearly as hard as what I’m asking for. We’re usually, like, a year ahead of what the clients end up asking for.
391 00:35:30.310 ⇒ 00:35:36.089 Uttam Kumaran: But it’s so funny, like, everybody on the AI team I brought on to help me build for our company.
392 00:35:36.240 ⇒ 00:35:44.140 Uttam Kumaran: we quickly found that other people need this, and now they’re all working on external stuff, so I’m back alone, for the most part.
393 00:35:44.900 ⇒ 00:35:46.609 Samprita H: That’s a good problem to have.
394 00:35:46.610 ⇒ 00:35:48.679 Uttam Kumaran: Yeah, yeah, yeah, yeah, yeah.
395 00:35:49.660 ⇒ 00:35:53.050 Uttam Kumaran: Okay, well, I appreciate it. Thank you so much. Yeah, I’ll talk to you soon.
396 00:35:53.880 ⇒ 00:35:56.049 Samprita H: Yeah, you too have a nice day. Bye.
397 00:35:56.050 ⇒ 00:35:56.970 Uttam Kumaran: Yeah, you too, bye.