Meeting Title: Brainforge Interview w- Awaish Date: 2026-02-20 Meeting participants: Lagani Patel, Awaish Kumar
WEBVTT
1 00:05:32.540 ⇒ 00:05:33.310 Lagani Patel: Hello.
2 00:05:34.410 ⇒ 00:05:34.980 Awaish Kumar: Bye.
3 00:05:35.270 ⇒ 00:05:36.120 Lagani Patel: Yay!
4 00:05:36.790 ⇒ 00:05:37.530 Awaish Kumar: How you doing?
5 00:05:37.780 ⇒ 00:05:39.319 Lagani Patel: I’m doing great, how are you?
6 00:05:40.430 ⇒ 00:05:43.670 Awaish Kumar: I’m great. Yeah, so…
7 00:05:43.810 ⇒ 00:05:52.840 Awaish Kumar: Yeah, my name is Avish Kumar, and this interview, we are… it is just a kind of intro call. We are…
8 00:05:52.980 ⇒ 00:06:00.310 Awaish Kumar: to talk a little bit about BrainForge, and then we are going to just briefly understand
9 00:06:00.440 ⇒ 00:06:03.440 Awaish Kumar: About your experience, and…
10 00:06:03.680 ⇒ 00:06:12.800 Awaish Kumar: And the past projects. So, I’m Avesh Kumar, and I’ve been working as a data engineer at, like…
11 00:06:12.970 ⇒ 00:06:19.919 Awaish Kumar: data engineering lead at Brainforge, and before that, I have kind of ground 10 years of experience working as a data engineer.
12 00:06:19.920 ⇒ 00:06:20.440 Lagani Patel: Nice.
13 00:06:21.610 ⇒ 00:06:22.890 Awaish Kumar: Indeed.
14 00:06:23.450 ⇒ 00:06:29.029 Awaish Kumar: BrainForge is a data and AI consultancy service where we…
15 00:06:29.600 ⇒ 00:06:35.970 Awaish Kumar: Normally provide, services to medium to large-scale enterprises.
16 00:06:36.760 ⇒ 00:06:46.059 Awaish Kumar: It operates remotely, like, it’s 100% remote. We have employees from across the world, so we have…
17 00:06:46.180 ⇒ 00:06:53.439 Awaish Kumar: We’re working from US, Europe, Asia, most of… but most of our clients are in US, so we…
18 00:06:54.000 ⇒ 00:07:08.180 Awaish Kumar: The… the way we… the way we communicate is mostly, like, we have some hours where we kind of overlap with our clients, with our colleagues, but then we are flexible to
19 00:07:08.340 ⇒ 00:07:14.210 Awaish Kumar: With people to work on their own hours. And normally,
20 00:07:14.860 ⇒ 00:07:21.029 Awaish Kumar: Whatever what happens in remote companies is that, you know, we,
21 00:07:21.130 ⇒ 00:07:25.709 Awaish Kumar: do a lot of async communication, writing documentation, communication, Slack.
22 00:07:26.010 ⇒ 00:07:33.240 Awaish Kumar: So, that’s how the… kind of the culture here. So, that’s basically it about Brainforge.
23 00:07:33.600 ⇒ 00:07:40.850 Awaish Kumar: So I would like to know about you now, you can introduce yourself with…
24 00:07:42.470 ⇒ 00:07:51.269 Lagani Patel: Yes, so hi, I’m Lagny. I have… I’m currently working as an AI software engineer at a startup called FinOptima Solutions.
25 00:07:51.270 ⇒ 00:08:06.650 Lagani Patel: I was hired to work on ML inference on a production level, and… but since it’s a startup, I also get to work on… work with the data that we are getting, and I work on the front end as well.
26 00:08:06.650 ⇒ 00:08:11.400 Lagani Patel: I… and I’m fairly new to the company, it’s been, like, maybe a…
27 00:08:11.400 ⇒ 00:08:21.880 Lagani Patel: month, approximately, now. And before that, I worked as a data scientist at the Museum of Natural Sciences in Raleigh, North Carolina.
28 00:08:21.940 ⇒ 00:08:32.959 Lagani Patel: There we were trying to estimate the mammal density across the United States, so I got a good exposure to, how to manage large-scale data there.
29 00:08:33.159 ⇒ 00:08:37.429 Lagani Patel: And outside of tech, I’ve worked as an event manager for 2 years.
30 00:08:37.530 ⇒ 00:08:43.340 Lagani Patel: And it has, sharpened my… a lot of… a lot… a lot of my soft skills.
31 00:08:44.450 ⇒ 00:08:50.229 Lagani Patel: I am good with communication, and I can think, quick on my feet.
32 00:08:51.870 ⇒ 00:09:00.710 Awaish Kumar: Okay, since you mentioned it’s just been a month or two working with your current company, why are you looking for a new job?
33 00:09:00.970 ⇒ 00:09:24.679 Lagani Patel: Yeah, so I am volunteering at the job right now. They are gonna pay me in some time, but since it’s a startup, I would also like to look for some sort of security. I definitely get to learn a lot, and I do want to be in a field where I get a challenging environment, which I definitely am getting here, but I also, being an adult, I feel like
34 00:09:24.910 ⇒ 00:09:27.899 Lagani Patel: I should always look for better opportunities.
35 00:09:29.140 ⇒ 00:09:40.710 Lagani Patel: And since, I think, I would also like to work on a remote, place, remote and in Austin, I think Brain Forge is based in Austin, right? I’m also based in Austin, so…
36 00:09:42.180 ⇒ 00:09:49.420 Awaish Kumar: Yeah, but it’s completely remote, like, our CEO is from Austin, but it’s just… Everybody works remotely.
37 00:09:49.750 ⇒ 00:09:53.930 Lagani Patel: Yeah, fair, that’s, that’s fair. It’s just, it might be easier to coordinate, maybe.
38 00:09:55.810 ⇒ 00:09:59.350 Awaish Kumar: So…
39 00:09:59.630 ⇒ 00:10:07.369 Awaish Kumar: In terms of your current role, what, like, what are you planning? Like, you’re planning to continue both the roles, or how… what are you thinking?
40 00:10:07.370 ⇒ 00:10:16.600 Lagani Patel: if I get a better opportunity, I think I might leave my company and join if I get another opportunity.
41 00:10:18.260 ⇒ 00:10:18.890 Lagani Patel: Yep.
42 00:10:18.890 ⇒ 00:10:20.340 Awaish Kumar: Okay, so…
43 00:10:20.540 ⇒ 00:10:35.590 Awaish Kumar: Since you mentioned you are very good at the communication skills, so I would like to get to more in details on how you normally communicate. So, if there are, like, you have worked in data science.
44 00:10:35.760 ⇒ 00:10:36.130 Lagani Patel: Yes.
45 00:10:36.130 ⇒ 00:10:43.700 Awaish Kumar: capacity. Also, so how do you communicate data uncertainties to the non-technical stakeholders?
46 00:10:43.880 ⇒ 00:10:46.170 Lagani Patel: Yes. So, to my benefit.
47 00:10:46.170 ⇒ 00:10:46.890 Awaish Kumar: Officially.
48 00:10:46.890 ⇒ 00:10:47.820 Lagani Patel: Sorry, yes?
49 00:10:47.820 ⇒ 00:10:51.819 Awaish Kumar: Especially if your result does not align with your expectation.
50 00:10:52.310 ⇒ 00:11:08.919 Lagani Patel: Yes, for sure. So, while working at the museum, I was working with a lot of, ecologists and, people with non-tech… proper non-tech background. What I would focus on is try… I would try to make data
51 00:11:08.920 ⇒ 00:11:22.940 Lagani Patel: easy to interpret. Like, even if I’m making a dashboard, I will make sure that it is in their terms, instead of just putting random technical terms that they might not understand. And I would try to simplify things as much as possible.
52 00:11:22.960 ⇒ 00:11:24.649 Lagani Patel: that I can…
53 00:11:24.770 ⇒ 00:11:49.079 Lagani Patel: For example, if an image is full of noise, I can’t just say that it is full of noise. I would put it in their terms, like the image is full of leaves, you can’t really see what is happening in that image. And other than that, what I did to make sure that we had a better understanding, I would constantly communicate what I am doing with the data.
54 00:11:49.080 ⇒ 00:12:12.219 Lagani Patel: or how I’m planning to handle it, and there was always a back and forth with my supervisor. He was very supportive, and I think just letting him know what I’m doing, why it is necessary, and what kind of results it will, we will get out of it. And yes, sometimes the expectations might not, might not be as…
55 00:12:12.310 ⇒ 00:12:29.789 Lagani Patel: might not be up to the standards, what they were expecting, but I think at the end of the day, with research, it is given that sometimes you might not get what you want, and I made sure that I communicate why we did not get what we wanted. So, we had a pretty good understanding in that way.
56 00:12:30.780 ⇒ 00:12:37.989 Awaish Kumar: Yeah. So… My point is more like, for example.
57 00:12:38.780 ⇒ 00:12:41.919 Awaish Kumar: have to communicate with your CEO.
58 00:12:41.920 ⇒ 00:12:43.040 Lagani Patel: Okay.
59 00:12:43.800 ⇒ 00:12:50.180 Awaish Kumar: And, and you… mentioned a few things, like, when starting a project, we
60 00:12:51.350 ⇒ 00:12:57.210 Awaish Kumar: Came up with some figures that, okay, we are going to optimize 5% of Oh, and…
61 00:12:57.970 ⇒ 00:13:04.430 Awaish Kumar: logistic cost or whatever. But using our, like, models.
62 00:13:04.640 ⇒ 00:13:14.479 Awaish Kumar: Or the… the, like, database decisions, but we didn’t reach that level. So how would you, like…
63 00:13:14.640 ⇒ 00:13:29.050 Awaish Kumar: How would you communicate that? Like, because now it is really impacting money, which is really a hard dis… that will be the hardest part of your communication to the non-technical stakeholder when it comes to money. So how would you…
64 00:13:29.450 ⇒ 00:13:32.589 Awaish Kumar: Make, how would you, like…
65 00:13:33.050 ⇒ 00:13:37.660 Awaish Kumar: Make it clear that, okay, what went wrong, or something like that.
66 00:13:38.660 ⇒ 00:13:44.810 Lagani Patel: Okay, so… For… I think I would focus on clarity.
67 00:13:44.930 ⇒ 00:13:50.560 Lagani Patel: like, first of all, before even deciding, I would scope
68 00:13:50.620 ⇒ 00:13:59.420 Lagani Patel: what they… what their expectations are, and what can actually be done. So I would… so…
69 00:13:59.480 ⇒ 00:14:14.700 Lagani Patel: let’s say if I’m, if I’m to optimize the, cost, right? So, when I… at the start of the project, I would make, clear definitions, like, what is the current baseline cost, or what exact metric are we gonna optimize.
70 00:14:14.700 ⇒ 00:14:24.789 Lagani Patel: And what are the trade-offs as well? Like, sometimes, you want, high speed, but it would also cost us more. So, I would clarify all the trade-offs.
71 00:14:24.820 ⇒ 00:14:28.500 Lagani Patel: trade-offs and their expectations on that. And…
72 00:14:28.700 ⇒ 00:14:48.519 Lagani Patel: So, and instead of, just saying that we could save a few here and there, I would make a detailed, research document or something that would, make them understand why we were not able to do what we were planning to do. And I would definitely try my best to come up with another plan.
73 00:14:48.710 ⇒ 00:14:51.889 Lagani Patel: Like, we could not do this, but we can definitely try this.
74 00:14:53.580 ⇒ 00:14:55.670 Lagani Patel: So, yeah.
75 00:14:56.610 ⇒ 00:15:02.769 Awaish Kumar: Okay, moving on… From… from that, too.
76 00:15:03.120 ⇒ 00:15:08.649 Awaish Kumar: For example… now, there’s one other scenario where, for example, we have a task which…
77 00:15:08.920 ⇒ 00:15:16.370 Awaish Kumar: Needs to be done, and you came up with a plan, and then how… or you can say.
78 00:15:16.600 ⇒ 00:15:23.399 Awaish Kumar: You came up with a, like, kind of a technical specification, how would you approach that?
79 00:15:23.840 ⇒ 00:15:29.630 Awaish Kumar: And now you are showcasing that to the team, and then there is a disagreement. Okay.
80 00:15:29.980 ⇒ 00:15:33.250 Awaish Kumar: In the team, how… on… on how to approach that.
81 00:15:33.630 ⇒ 00:15:40.759 Awaish Kumar: So now… that’s, like, solution. So how… how would you resolve those, disagreements in the…
82 00:15:40.760 ⇒ 00:15:45.619 Lagani Patel: Yeah. Right. So, resolving conflicts,
83 00:15:46.180 ⇒ 00:15:52.610 Lagani Patel: Yes, can I just gather my thoughts for, on this for a moment? I think…
84 00:15:55.180 ⇒ 00:16:04.089 Lagani Patel: Yes. So, for example, this one time at the museum, we… we had a problem where we were…
85 00:16:04.450 ⇒ 00:16:19.779 Lagani Patel: we were using a software which was costing us 200 per month and use that money somewhere else.
86 00:16:19.780 ⇒ 00:16:27.739 Lagani Patel: So, I could not… I cannot expect anybody to just accept what I’m saying, and, just…
87 00:16:27.920 ⇒ 00:16:42.169 Lagani Patel: be okay with it. So what I did was, I made sure that I understand what I’m saying, first of all, fully. Like, I would give it a proper thought, I would be objective, I would look at the constraints or the timeline.
88 00:16:42.340 ⇒ 00:16:58.170 Lagani Patel: or anything that is related to that decision, I would… and since it is about cost cutting, I would, make sure that where… if we are seeing… saving some money there, and where we can spend it better, I would definitely come up with a proper plan of
89 00:16:58.170 ⇒ 00:17:14.880 Lagani Patel: So that would be a migration, right? So I would come up with a plan on how to migrate everything, how to make things easier for the migration, then why we should be doing that, and how it is going to impact our, further, our process further.
90 00:17:14.990 ⇒ 00:17:32.699 Lagani Patel: So, and if somebody disagrees, like, somebody did disagree with me on that, I will try to convince them, but if they have a very good and valid point, I would disagree and still commit to the process. At the end of the day, we all share a common goal.
91 00:17:32.970 ⇒ 00:17:34.949 Lagani Patel: We all want something better for the team.
92 00:17:34.950 ⇒ 00:17:45.080 Awaish Kumar: What do you do? That’s the point, that somebody disagreed. Did you just left what you were doing, or how did you convince them?
93 00:17:45.480 ⇒ 00:17:58.970 Lagani Patel: Yes, so I came up with a very detailed documentation on why we should do it. I came up with very clear reasons, like how it would benefit us, the project, and the overall lab itself.
94 00:17:58.970 ⇒ 00:18:16.299 Lagani Patel: So, even if they disagreed, because we had another deadline at the moment, I understood that what they are saying is also right, and I don’t… I would not mind being wrong in any place. If they also have a valid point, I would completely understand.
95 00:18:16.330 ⇒ 00:18:18.889 Lagani Patel: And try to meet at a common ground.
96 00:18:21.650 ⇒ 00:18:27.480 Awaish Kumar: Okay, then I think it’s an interesting, situation, like,
97 00:18:27.850 ⇒ 00:18:33.939 Awaish Kumar: We are just taking that project as an example now, that you mentioned that.
98 00:18:36.510 ⇒ 00:18:52.549 Awaish Kumar: that you propose a new tool to save some cost, so I would like to know, like, is… is migration of a tool where you require learning and a migration plan?
99 00:18:52.790 ⇒ 00:18:58.870 Awaish Kumar: is, Is, is, what’d you say?
100 00:18:59.450 ⇒ 00:19:00.030 Lagani Patel: Worth it?
101 00:19:00.030 ⇒ 00:19:05.110 Awaish Kumar: It’s value. It’s worth it when it’s just $200 per month.
102 00:19:06.260 ⇒ 00:19:25.990 Lagani Patel: So I think that really depends on what kind of company I’m working for. Since we were a research team, and we, at that time, grants were, not… we were working on a grant, and we would like… we would like to optimize our cost as much as possible. That was a geospatial analysis tool.
103 00:19:26.200 ⇒ 00:19:41.319 Lagani Patel: Which we did use quite often, but I would… but I also realized that that money could have been saved, and we could put it in GPU usage, or, getting something better for the lab. So, for that company… for that lab.
104 00:19:41.470 ⇒ 00:19:58.039 Lagani Patel: cost was one of the concerns, because we were… we were very limited on that. But for a company, for example, they have good, good funding or something like that. I… if their cost is not their main concern, then maybe, yes, migration would not make much sense.
105 00:19:59.100 ⇒ 00:20:05.310 Awaish Kumar: Yeah, it takes, like, the… The migration takes time, it requires resources.
106 00:20:05.310 ⇒ 00:20:05.800 Lagani Patel: Yeah.
107 00:20:05.810 ⇒ 00:20:10.690 Awaish Kumar: Also, like, we are spending money on migration, so…
108 00:20:11.940 ⇒ 00:20:27.350 Lagani Patel: Yeah, sorry. So that was a geo, spatial analysis tool. It did require some setting or some setup. Since they were non-technical people, for them, it was harder than what, what a technical per… how…
109 00:20:27.350 ⇒ 00:20:42.260 Lagani Patel: See, MATLAB might be easier for me, but somebody who’s never used it, like, somebody… ecologist or somebody like that, it might be hard for them. So, what I… the migration plan would be just them
110 00:20:42.600 ⇒ 00:20:44.600 Lagani Patel: You know?
111 00:20:45.460 ⇒ 00:20:56.880 Lagani Patel: the migration would be just teaching them how to actually use the tool, and just setting up our database directly to that, connect that to the database. It was not.
112 00:20:56.880 ⇒ 00:20:57.560 Awaish Kumar: Okay.
113 00:20:57.560 ⇒ 00:20:58.130 Lagani Patel: Yep.
114 00:20:58.630 ⇒ 00:21:03.500 Awaish Kumar: Moving on to… Not…
115 00:21:04.260 ⇒ 00:21:10.030 Awaish Kumar: If you’ve given me an example of any one of your recent projects where you worked
116 00:21:10.280 ⇒ 00:21:13.370 Awaish Kumar: Like, end-to-end. And,
117 00:21:13.490 ⇒ 00:21:21.079 Awaish Kumar: Yeah, just… just walk me through it, how you did it, what you did, and what tools and technologies were used, and things like that.
118 00:21:21.420 ⇒ 00:21:32.429 Lagani Patel: Yes, so my most recent project would be a voice AI agent that I built called Shanti. I made it, because
119 00:21:32.430 ⇒ 00:21:48.969 Lagani Patel: I, I would defi- I like to use AI for things that I’m actually gonna use… I actually need. For example, I used… I made that, AI agent, where it would talk to you whenever you’re feeling anxious about something, or you need to just,
120 00:21:49.110 ⇒ 00:21:55.440 Lagani Patel: You need a clear thinking, or you need… you just need somebody to talk to.
121 00:21:55.860 ⇒ 00:22:07.429 Lagani Patel: It is… it was no… in no way or form, it was a, therapist, but it is just a real, real, real-time voice-to-voice conversation AI bot, or AI, agent, sorry.
122 00:22:07.760 ⇒ 00:22:12.779 Lagani Patel: So yes, I… how I built it. I built it on… I… sorry.
123 00:22:13.700 ⇒ 00:22:17.590 Awaish Kumar: Yeah, was it a project where you were doing in a company, or is it just a side?
124 00:22:17.590 ⇒ 00:22:20.559 Lagani Patel: No, it’s a side project, I just did it for myself.
125 00:22:21.020 ⇒ 00:22:27.270 Awaish Kumar: Okay, can we talk about something which you’ve worked on in one of your jobs?
126 00:22:28.220 ⇒ 00:22:29.590 Lagani Patel: At my job.
127 00:22:30.480 ⇒ 00:22:31.160 Awaish Kumar: Yeah, yeah.
128 00:22:31.290 ⇒ 00:22:40.149 Lagani Patel: Yeah, okay. So currently, I’m working on develop, on developing that, like, a via… so at,
129 00:22:40.570 ⇒ 00:22:57.869 Lagani Patel: FinOptima, we are developing Audio Voice Guard for fraud detection. And we already have a behavior analysis model in place, and a few other ML models, which will be scoring how risky the call is, or if fraud is detected in that or not.
130 00:22:58.200 ⇒ 00:23:13.530 Lagani Patel: But, yes, so what I’m doing is, right now, I’m building schemas, and I’m handling all the incoming data. For example, if a call is, if you’re receiving a call, you will have some metadata around it as well. So…
131 00:23:13.650 ⇒ 00:23:25.220 Lagani Patel: I’m working on developing schemas, and I handle missing values around it, I validate the data, and after that, it goes to the ML…
132 00:23:25.220 ⇒ 00:23:26.300 Awaish Kumar: your source.
133 00:23:26.540 ⇒ 00:23:27.090 Lagani Patel: Sorry?
134 00:23:27.090 ⇒ 00:23:28.720 Awaish Kumar: What is the source of the data?
135 00:23:29.070 ⇒ 00:23:41.590 Lagani Patel: If somebody is calling… right now, we are still, developing demo, so, it is just gonna… for example, you are call… calling somebody, and your, voice call.
136 00:23:41.600 ⇒ 00:23:51.689 Lagani Patel: The audio of the call will be our source data, and the metadata around it, like the phone number, or the time… timestamp, or whoever it’s calling, yes.
137 00:23:51.690 ⇒ 00:24:01.159 Awaish Kumar: I mean, is that call happening on your platform? Is it in an app, or is it a… It would be a normal… Yeah.
138 00:24:01.160 ⇒ 00:24:08.280 Lagani Patel: So, it would be a normal phone call. We are connecting it to our pipeline using, I think, AWS Connect.
139 00:24:08.380 ⇒ 00:24:15.019 Lagani Patel: if I’m not wrong, you can use that to connect a live phone call to your… to a pipeline.
140 00:24:15.190 ⇒ 00:24:17.200 Lagani Patel: That would be our incoming data.
141 00:24:18.220 ⇒ 00:24:19.010 Awaish Kumar: Okay.
142 00:24:19.230 ⇒ 00:24:19.730 Lagani Patel: Yeah.
143 00:24:21.180 ⇒ 00:24:21.880 Lagani Patel: And…
144 00:24:21.880 ⇒ 00:24:23.029 Awaish Kumar: Okay, don’t delay.
145 00:24:23.210 ⇒ 00:24:41.610 Lagani Patel: Yeah, so yes, after that, I’m also, yes, I have to get that data ready for the, ML… for ML inference. I, yes, I validate the data, I handle missing values or null values, and we also divide the, audio into, chunks of 2 to 3 seconds.
146 00:24:42.450 ⇒ 00:24:56.249 Lagani Patel: Before putting that… all of that data in the ML model, we, go through… so we make a package. We… the package will have temporal, temporal chunks. Like, for example, if this
147 00:24:56.250 ⇒ 00:25:15.679 Lagani Patel: during the phone call, there is a fraud detected in this part, but overall, the call was, call was fine. So, we would also need context if, if it… if we do just bit by bit, which is stateless, it might not make sense. So, we would make a whole package around the metadata and the audio clips.
148 00:25:16.060 ⇒ 00:25:24.199 Lagani Patel: And, that the whole package will be scheduled to go, to the ML model, using Redis, Redis Q.
149 00:25:25.870 ⇒ 00:25:27.960 Awaish Kumar: Can we talk more, a little bit about
150 00:25:28.420 ⇒ 00:25:31.920 Awaish Kumar: like, tools and technologies, like, for example, you mentioned
151 00:25:32.420 ⇒ 00:25:42.740 Awaish Kumar: the data source is the call, and you’re using AWS Connect to get that data, but where it goes, like, does that… where that data lives? What is the…
152 00:25:43.020 ⇒ 00:25:51.510 Awaish Kumar: Storage, what… how the transformation happens, what tools do you use, what scripts do you write, then how…
153 00:25:52.070 ⇒ 00:25:55.539 Awaish Kumar: How they are orchestrated, and then finally.
154 00:25:55.740 ⇒ 00:26:04.170 Awaish Kumar: Like, what tools to use to, for example, Validated, or…
155 00:26:04.500 ⇒ 00:26:09.389 Awaish Kumar: Yeah, like, how model basically interacts with your data.
156 00:26:09.790 ⇒ 00:26:21.880 Lagani Patel: Yeah, so we are using, DynamoDB right now, because we wanted more, more of a flexible, data storage, so we are using DynamoDB for,
157 00:26:22.110 ⇒ 00:26:26.200 Lagani Patel: metadata, and we are, I think, using Lambda…
158 00:26:26.200 ⇒ 00:26:44.279 Lagani Patel: not sorry, not Lambda, I think S3 bucket for our audio call, audio clips. And I am majorly using Python for, all the… since you can’t really do schema, you can’t build schemas in DynamoDB itself, I’m using Python to define everything.
159 00:26:44.280 ⇒ 00:27:00.600 Lagani Patel: And we were actually considering AW, sorry, Airflow, but what happened is, it is a… this is more like a live, incoming call and a real-time inference. So, and Airflow is, I feel like, better for a scheduled job.
160 00:27:01.770 ⇒ 00:27:05.430 Lagani Patel: So, right now, we are orchestrating using Python scripts.
161 00:27:05.980 ⇒ 00:27:12.840 Lagani Patel: But we are also looking into solutions where we can use some, a third-party tool for the orchestration.
162 00:27:13.720 ⇒ 00:27:18.280 Awaish Kumar: Like, maybe Python scripts, where they are running.
163 00:27:18.800 ⇒ 00:27:22.790 Lagani Patel: They are running on, cloud, AWS cloud.
164 00:27:23.680 ⇒ 00:27:31.190 Awaish Kumar: I mean… Like, can you, like, a little bit talk about how they are running, like…
165 00:27:31.890 ⇒ 00:27:36.470 Lagani Patel: Okay, so, we have… I think…
166 00:27:36.670 ⇒ 00:27:41.780 Lagani Patel: We have S3 buckets in place, and I think we’re using SageMaker for the ML inference.
167 00:27:43.640 ⇒ 00:27:56.199 Lagani Patel: And for Pi… the Python scripts are running… I think they are also… I think they are in EC… ECS containers on, AWS.
168 00:27:58.050 ⇒ 00:27:58.860 Awaish Kumar: Okay.
169 00:27:59.800 ⇒ 00:28:04.789 Awaish Kumar: Hmm… in the, like, ECS containers are just containers where you can.
170 00:28:04.790 ⇒ 00:28:05.370 Lagani Patel: Yeah.
171 00:28:05.370 ⇒ 00:28:07.630 Awaish Kumar: For your images, like, how you…
172 00:28:08.280 ⇒ 00:28:11.960 Awaish Kumar: How you run it, like, how you… like, the… where they are running.
173 00:28:13.780 ⇒ 00:28:14.440 Awaish Kumar: execution?
174 00:28:16.440 ⇒ 00:28:25.550 Lagani Patel: So, I think, we’re using one of the AWS services, let me just think, let me just try and remember the name of that.
175 00:28:25.850 ⇒ 00:28:30.929 Lagani Patel: service… I think Lambda? No, AWS Lambda was…
176 00:28:35.620 ⇒ 00:28:45.149 Awaish Kumar: Okay, we can just move on. So, for example, you mentioned the data comes in DynamoDB, then you have audio clips in S3.
177 00:28:45.160 ⇒ 00:28:46.190 Lagani Patel: So, is…
178 00:28:46.230 ⇒ 00:28:50.189 Awaish Kumar: Is that really our audio clips you’re interacting with, or…
179 00:28:51.780 ⇒ 00:28:54.380 Awaish Kumar: Or is it something, like, something else?
180 00:28:55.220 ⇒ 00:28:57.249 Lagani Patel: We are handling audio clips.
181 00:28:57.930 ⇒ 00:29:02.539 Awaish Kumar: I mean, like… how you are getting features from AutoClip.
182 00:29:03.180 ⇒ 00:29:16.699 Lagani Patel: So, we have a behavior analysis model in place. That behavior analysis model will analyze the audio and keep giving its scores, like, if they detect any kind of fraud in it.
183 00:29:18.340 ⇒ 00:29:23.129 Awaish Kumar: So that… that… behavioral model, where, like, is it your…
184 00:29:23.480 ⇒ 00:29:29.000 Awaish Kumar: Is it your model? Is it, like, some kind of tool which takes on audio clip and gives some responses, or…
185 00:29:29.000 ⇒ 00:29:50.550 Lagani Patel: I think it’s a model. I have not developed it, so I might not be able to get into more details around the behavioral model, but I do know better about the ML… so that is behavioral analysis, and after that, it goes to our ML model. ML model, we have built using two different layers. It is… it has XLS model, and.
186 00:29:50.550 ⇒ 00:29:56.749 Awaish Kumar: using… Are you using Aren’t you using transcripts of the call?
187 00:29:57.760 ⇒ 00:29:59.440 Lagani Patel: sorry, JavaScript?
188 00:29:59.870 ⇒ 00:30:09.630 Lagani Patel: Transcripts. Transcripts. I think we are, but, just with transcript, it might not be… we might not be able to detect,
189 00:30:09.830 ⇒ 00:30:24.849 Lagani Patel: So, fraud, basically what our model is doing, we are not exactly using the transcript, we are using the depth of the voice, or we are detecting if, the voice sounds very robotic, or if they are nervous about something.
190 00:30:25.650 ⇒ 00:30:30.430 Lagani Patel: So we are… we’ll have to use the actual audio instead of the transcript.
191 00:30:31.100 ⇒ 00:30:36.299 Awaish Kumar: And then that data… a big, like…
192 00:30:36.580 ⇒ 00:30:40.549 Awaish Kumar: pick… go somewhere, like, so where that lives, like…
193 00:30:41.700 ⇒ 00:30:47.359 Lagani Patel: So that data, goes through… it is stored in S3 bucket.
194 00:30:48.100 ⇒ 00:30:50.180 Awaish Kumar: Okay, that also goes to S3.
195 00:30:50.180 ⇒ 00:30:50.930 Lagani Patel: Yes.
196 00:30:51.210 ⇒ 00:30:55.870 Awaish Kumar: Okay, we don’t… I’m trying to understand, when you are building the schema and stuff.
197 00:30:56.360 ⇒ 00:31:01.119 Awaish Kumar: So is it still in the stream? Like, you’re not storing in any database?
198 00:31:01.740 ⇒ 00:31:14.580 Lagani Patel: we’re using DynamoDB to store the metadata, but since DynamoDB is NoSQL, we won’t be able to write down schemas or table in DynamoDB directly.
199 00:31:15.490 ⇒ 00:31:17.999 Lagani Patel: But we would also… yeah.
200 00:31:20.090 ⇒ 00:31:23.769 Awaish Kumar: Okay, we have DynamoDB, where you’re restoring your documents, then we have
201 00:31:24.010 ⇒ 00:31:26.919 Awaish Kumar: S3, where you are… you have audio clips, and then…
202 00:31:27.130 ⇒ 00:31:38.580 Awaish Kumar: you ran through some model, and then you got some features. I’m asking, what is the storage for those features? Is it just S3? Is it some other database?
203 00:31:38.980 ⇒ 00:31:55.260 Lagani Patel: So the, I think, I think it’s gonna be DynamoDB right now. Whatever results, for example, we are getting score, risk, risk score, or we are writing a rational explanation why, why the call were…
204 00:31:55.280 ⇒ 00:32:02.970 Lagani Patel: call was declared fraud or not. So we have, so all of that details are right now in DynamoDB.
205 00:32:03.980 ⇒ 00:32:09.700 Awaish Kumar: Okay, and then, your model, the final model that you’re working on.
206 00:32:10.140 ⇒ 00:32:14.740 Awaish Kumar: What does… what… like, what you do with… in that model?
207 00:32:15.520 ⇒ 00:32:29.630 Lagani Patel: Yes, so that model… that ML model, it has, two different, it is a two-layered architecture. So, the first one… the first model, XLS, it detects the, highness or,
208 00:32:30.220 ⇒ 00:32:41.540 Lagani Patel: more around heuristics around the voice and the audio, and there is another model, a neural network model. I think we are using CNN, or…
209 00:32:42.370 ⇒ 00:32:50.480 Lagani Patel: most probably it is CNN. And we use that model to understand the base of the audio call.
210 00:32:51.560 ⇒ 00:32:56.849 Awaish Kumar: Okay, so, like, are you the one actually working on this? No, I…
211 00:32:57.130 ⇒ 00:33:07.049 Lagani Patel: Yeah, so it is already in progress, and it is already in production, most probably. I am working into setting up the pipeline.
212 00:33:08.310 ⇒ 00:33:21.209 Lagani Patel: of how the audio will interact with the ML model, what kind… so, instead of just putting the… whatever data that we’re getting directly in the ML model, I’m working on cleaning the data, validating it, standardizing it.
213 00:33:21.590 ⇒ 00:33:27.849 Lagani Patel: And creating a whole package around it, and then feed that package to the ML model.
214 00:33:29.290 ⇒ 00:33:30.290 Awaish Kumar: Okay.
215 00:33:30.290 ⇒ 00:33:30.640 Lagani Patel: Yeah.
216 00:33:31.190 ⇒ 00:33:31.940 Lagani Patel: I’m sorry.
217 00:33:31.940 ⇒ 00:33:36.119 Awaish Kumar: So, what are you using to clean up and all… to do all those things?
218 00:33:36.120 ⇒ 00:33:38.130 Lagani Patel: Python. Right now, I’m using Python.
219 00:33:38.810 ⇒ 00:33:43.110 Awaish Kumar: okay, you’re not using Python, but then you are not deploying those Python scripts?
220 00:33:43.980 ⇒ 00:33:46.480 Lagani Patel: I am deploying…
221 00:33:48.190 ⇒ 00:33:56.280 Lagani Patel: So, right now, I think it’s still… it’s very primitive, and we are still working on it, so it’s not completely deployed, deployed yet.
222 00:33:56.550 ⇒ 00:33:57.410 Awaish Kumar: Huh.
223 00:33:57.930 ⇒ 00:33:58.600 Lagani Patel: Yeah.
224 00:33:58.600 ⇒ 00:33:59.240 Awaish Kumar: Okay.
225 00:33:59.430 ⇒ 00:34:05.660 Awaish Kumar: Like, is there any… like…
226 00:34:05.840 ⇒ 00:34:12.610 Awaish Kumar: Is there any other project, like, where you have deployed, something? You can just briefly talk about that?
227 00:34:13.710 ⇒ 00:34:17.809 Lagani Patel: I’ve, deployed in the sense, on cloud.
228 00:34:18.989 ⇒ 00:34:27.979 Awaish Kumar: like, something which runs in production. I just want to know how… how would you run your system in production without you interacting manually with it?
229 00:34:29.570 ⇒ 00:34:33.250 Lagani Patel: Okay. Okay,
230 00:34:34.900 ⇒ 00:34:53.759 Lagani Patel: Yes, so, at the museum, while I was working, we, there also, I was running, I wrote Python scripts and automated a lot of stuff. There, we were running the whole, pipeline on a GPU, NVIDIA GPU. I think we were using 140…
231 00:34:54.239 ⇒ 00:35:03.869 Lagani Patel: 1409, or… it was a GPU, and we were using local storage since we wanted a higher computation power. So…
232 00:35:04.070 ⇒ 00:35:11.899 Lagani Patel: There I def… there I worked on more of a production level, a very large-scale database.
233 00:35:12.290 ⇒ 00:35:16.190 Lagani Patel: we had almost 100K images coming in every week.
234 00:35:19.210 ⇒ 00:35:24.209 Awaish Kumar: Okay, so how… Like, how those images were coming.
235 00:35:24.490 ⇒ 00:35:28.099 Awaish Kumar: In your database, how were you, basically.
236 00:35:28.270 ⇒ 00:35:34.000 Awaish Kumar: running your scripts, like, what would your info look like? That’s what I want to understand.
237 00:35:34.210 ⇒ 00:35:47.379 Lagani Patel: Yes, so we measurely had 3 different kind of categories of data. The first one would be image metadata, then the second one would be image… images itself, and we also had camera meta metadata.
238 00:35:47.810 ⇒ 00:36:07.639 Lagani Patel: So the first… the first ingestion job would be, all the re… a lot of different universities were participating in our study, so we… we had a common website called Wildlife Insights. They would upload their data on that, and I wrote a Python script, which would run locally on my GPU,
239 00:36:07.640 ⇒ 00:36:12.909 Lagani Patel: And that… that would extract all the meta, metadata from that
240 00:36:12.910 ⇒ 00:36:18.939 Lagani Patel: pipe, from that website, and it would… it would have corresponding image URLs, which would.
241 00:36:18.940 ⇒ 00:36:27.030 Awaish Kumar: Isn’t that… Was that an external website, or your internal portal, where the Indeed.
242 00:36:27.030 ⇒ 00:36:35.639 Lagani Patel: No, it is an external website called Wildlife Insights. It is used for research and research databases.
243 00:36:36.610 ⇒ 00:36:42.780 Awaish Kumar: And how would you… how were you getting data from them? Like, what was the… Way to extract data.
244 00:36:43.440 ⇒ 00:36:51.850 Lagani Patel: So you can just… I wrote a… I automated the thing by just writing a script where it would just download that data.
245 00:36:52.720 ⇒ 00:37:01.529 Lagani Patel: And it would unzip it. You will get a zip file, and you, you will, you can unzip it, and then you will,
246 00:37:02.210 ⇒ 00:37:06.710 Lagani Patel: then the Python script will go through the whole data set.
247 00:37:06.850 ⇒ 00:37:14.320 Lagani Patel: And run the basic validation steps, like missing values, null, checking nulls, outliers, and everything.
248 00:37:15.260 ⇒ 00:37:16.450 Lagani Patel: Okay.
249 00:37:16.450 ⇒ 00:37:21.430 Awaish Kumar: I think we are over the line, over the time, so…
250 00:37:21.800 ⇒ 00:37:27.869 Awaish Kumar: I don’t have any more questions, so yeah, if you have any questions, yeah, you can ask.
251 00:37:27.870 ⇒ 00:37:41.199 Lagani Patel: Yeah, for sure. So I wanted to know, how, Brainforge is, so are you guys… some of the companies don’t really prefer using AI for coding, but what is your stand on it?
252 00:37:43.150 ⇒ 00:37:47.560 Awaish Kumar: Yeah, like, Print Forge is an AI-powered startup.
253 00:37:47.920 ⇒ 00:37:50.110 Awaish Kumar: Everybody uses AI here.
254 00:37:50.110 ⇒ 00:37:50.450 Lagani Patel: Okay.
255 00:37:50.450 ⇒ 00:37:51.929 Awaish Kumar: from…
256 00:37:52.170 ⇒ 00:38:04.449 Awaish Kumar: Not just the engineers, not just the developers, but everybody here, and it doesn’t matter which department they belong, like, sales, marketing,
257 00:38:05.450 ⇒ 00:38:11.439 Awaish Kumar: And, other departments, like analysts, everybody uses AI to.
258 00:38:11.440 ⇒ 00:38:12.239 Lagani Patel: Okay, that’s nice.
259 00:38:12.560 ⇒ 00:38:16.920 Awaish Kumar: To improve their… to actually accelerate their… work.
260 00:38:17.650 ⇒ 00:38:27.479 Lagani Patel: Right. And how does this, team… how are the teams structured, here? Like, is everyone working together, or there are specific teams for specific tasks?
261 00:38:29.250 ⇒ 00:38:31.669 Awaish Kumar: So, like, we have a specialized…
262 00:38:32.300 ⇒ 00:38:41.860 Awaish Kumar: teams. We have a team with AI capabilities, we have a data engineering team, we have a data and analytics engineers, we have a
263 00:38:42.110 ⇒ 00:38:44.210 Awaish Kumar: And a team of analysts.
264 00:38:44.390 ⇒ 00:38:45.340 Awaish Kumar: So…
265 00:38:45.550 ⇒ 00:38:57.990 Awaish Kumar: And we have different leads for each team. We have data… I’m kind of DEA lead here, and we have a lead analyst lead, we have a…
266 00:38:58.370 ⇒ 00:39:06.630 Awaish Kumar: lead for AI team, we also have a lead for… Product analytics. So, basically.
267 00:39:06.840 ⇒ 00:39:16.059 Awaish Kumar: Kind of… this is how we are structured in terms of capabilities, but then how we work for the client is completely different.
268 00:39:16.150 ⇒ 00:39:19.669 Lagani Patel: Okay. Because for a single client, we don’t just need.
269 00:39:19.870 ⇒ 00:39:23.319 Awaish Kumar: all the data engineers on that, right? We just… we need a…
270 00:39:23.710 ⇒ 00:39:32.209 Awaish Kumar: team of people with different skill sets on a client. So, maybe a client port looks different than
271 00:39:32.280 ⇒ 00:39:43.040 Awaish Kumar: skill-based departments we have. A client port could be, like, mix of analyst, data analyst, data engineer, product analytics engineer, martech engineer.
272 00:39:43.090 ⇒ 00:39:53.509 Awaish Kumar: So, all they combine together to become a client port, and it completely depends on the client what kind of services they want from us, if they just need
273 00:39:53.700 ⇒ 00:39:59.369 Awaish Kumar: Data engineering services, then we will have just one or two data engineers, if they need
274 00:39:59.700 ⇒ 00:40:17.860 Awaish Kumar: more, like, from… not just data engineering services, but they also need someone to analyze it, and give them the insights, so we have a… we can bring in, like, the data analysts. Same goes for all other
275 00:40:18.100 ⇒ 00:40:23.870 Awaish Kumar: like, the… Team, like, their team of expertise, so, like.
276 00:40:24.090 ⇒ 00:40:26.439 Awaish Kumar: It depends on what client needs.
277 00:40:27.230 ⇒ 00:40:28.080 Lagani Patel: Right.
278 00:40:28.280 ⇒ 00:40:40.409 Lagani Patel: And, do you guys find it a little challenging, or is it going very smooth? Like, you have different people in different, continents, so what are the challenges do you guys usually face?
279 00:40:42.980 ⇒ 00:40:51.030 Awaish Kumar: The only challenge is, communication, obviously. Like, when you are working remote, if you don’t communicate,
280 00:40:51.390 ⇒ 00:41:04.519 Awaish Kumar: then it would be a… it could be a problem, because no… nobody knows what you are doing until you inform or you update. That’s what our,
281 00:41:04.680 ⇒ 00:41:10.230 Awaish Kumar: That’s what is in our process, that… we…
282 00:41:10.770 ⇒ 00:41:17.339 Awaish Kumar: like, it’s kind of part of culture here at Bridge Forge. We want… we encourage everybody to over-communicate.
283 00:41:17.340 ⇒ 00:41:17.870 Lagani Patel: Right.
284 00:41:18.920 ⇒ 00:41:23.809 Awaish Kumar: Write down select messages, write down documentation, write on linear, write on…
285 00:41:24.020 ⇒ 00:41:41.190 Awaish Kumar: Every… everywhere where you could, like, just, give your updates, right? Sometimes you just write in linear, nobody sees it. So instead of just, focusing on your parts, let’s try to just be over-communicative.
286 00:41:41.850 ⇒ 00:41:50.970 Awaish Kumar: maybe I can write in linear ticket, but I can also just go in Slack and tell my team, okay, I’ve worked on this, and etc.
287 00:41:51.210 ⇒ 00:42:04.049 Awaish Kumar: this is one of the things, and secondly, we have our, obviously, meeting set up. We have a setup every day when we meet, talk about clients’ health, and what the work we are doing, what we are…
288 00:42:04.150 ⇒ 00:42:15.239 Awaish Kumar: trying to achieve today, and then, at the end of the day, normally, people just communicate in the Slack channel that I was able to finish this, and not this, and that.
289 00:42:15.590 ⇒ 00:42:16.939 Awaish Kumar: And we have…
290 00:42:17.420 ⇒ 00:42:26.439 Awaish Kumar: it’s kind of weekly updates for the client as well, with what we achieved, what we delivered this week, and things like that. So, yeah.
291 00:42:27.700 ⇒ 00:42:33.189 Lagani Patel: I think that makes a lot of sense. Over-communication is definitely better than not communicating at all.
292 00:42:33.950 ⇒ 00:42:34.580 Awaish Kumar: Nap.
293 00:42:34.800 ⇒ 00:42:35.710 Lagani Patel: Makes sense.
294 00:42:36.670 ⇒ 00:42:38.790 Lagani Patel: Okay.
295 00:42:39.480 ⇒ 00:42:43.069 Lagani Patel: Sorry, I’m sorry if I’m taking too much of your time.
296 00:42:43.980 ⇒ 00:42:48.620 Awaish Kumar: No worries. If you have any other questions, you can ask. We still have, like, maybe 2-3 minutes.
297 00:42:49.060 ⇒ 00:42:57.760 Lagani Patel: Oh, okay. I wanted to know, like, is it fast-paced, or do you guys, I mean…
298 00:42:57.760 ⇒ 00:42:58.350 Awaish Kumar: Don’t use it.
299 00:42:58.350 ⇒ 00:42:59.770 Lagani Patel: The book is done, or…
300 00:43:01.120 ⇒ 00:43:15.379 Awaish Kumar: Yeah, as I mentioned, it’s a startup, AI-powered startup. We’re using AI to accelerate, our speed of delivery, so that means, it’s just, like, it’s a fast-paced environment, everybody will…
301 00:43:15.810 ⇒ 00:43:21.170 Awaish Kumar: Is delivering now, like, more than what they could have done maybe one year ago.
302 00:43:21.430 ⇒ 00:43:25.780 Awaish Kumar: With the use of AI. So, yeah, that’s how it is.
303 00:43:26.600 ⇒ 00:43:28.350 Lagani Patel: Yeah, that makes a lot of sense.
304 00:43:29.860 ⇒ 00:43:32.520 Lagani Patel: I think I’m good with my questions.
305 00:43:32.520 ⇒ 00:43:35.169 Awaish Kumar: I think then,
306 00:43:35.620 ⇒ 00:43:49.770 Awaish Kumar: like, yeah, thank you for your time, for the interview today. Rico from our operations team will reach out to you, maybe in a week’s time, with the next steps, and yeah, that’s it.
307 00:43:50.200 ⇒ 00:43:53.690 Lagani Patel: Thank you, thank you so much for having me today, I… this was really fun.
308 00:43:54.470 ⇒ 00:43:55.040 Awaish Kumar: Yep.
309 00:43:55.690 ⇒ 00:43:56.560 Lagani Patel: Living.