WEBVTT 00:00:00.000 --> 00:00:02.720 Hey, I'm Mukundan. This show is about solving 00:00:02.720 --> 00:00:06.080 real problems with small, useful AI, things that 00:00:06.080 --> 00:00:09.220 you can actually use today. Each week, we pick 00:00:09.220 --> 00:00:12.279 one problem, build a simple workflow or tool, 00:00:12.400 --> 00:00:14.820 and talk through the decisions, what to automate, 00:00:15.060 --> 00:00:18.359 how to check quality, and how to make it reliable. 00:00:19.379 --> 00:00:22.760 If you're a builder, creator, or just AI curious, 00:00:23.120 --> 00:00:25.899 you'll leave with the steps that you can copy 00:00:25.899 --> 00:00:46.350 tonight. You know those quiet moments before 00:00:46.350 --> 00:00:49.210 a project begins, when the cursor blinks on a 00:00:49.210 --> 00:00:52.329 blank screen and you just sit there, waiting 00:00:52.329 --> 00:00:56.429 for the spark? That's how this all started. No 00:00:56.429 --> 00:01:00.350 big idea. No vision board. Just curiosity and 00:01:00.350 --> 00:01:04.250 a little frustration. I was looking at a dataset 00:01:04.250 --> 00:01:08.590 at work. A big one. Customer transactions across 00:01:08.590 --> 00:01:12.709 multiple months. And I realized something that... 00:01:12.810 --> 00:01:15.930 stopped me in my tracks. I didn't know what to 00:01:15.930 --> 00:01:20.370 ask. Sure, I could slice it by week, by product, 00:01:20.450 --> 00:01:25.810 by region as well. But that wasn't insight. That 00:01:25.810 --> 00:01:32.730 was just navigation. And it hit me. If I, someone 00:01:32.730 --> 00:01:37.230 who lives and breathes analytics, sometimes struggle 00:01:37.230 --> 00:01:40.049 to know where to begin, what chance do others 00:01:40.049 --> 00:01:41.870 have? Like people on the business side maybe, 00:01:41.989 --> 00:01:45.849 right? And that moment became the seed of an 00:01:45.849 --> 00:01:50.290 idea. Could I build an AI that thinks like an 00:01:50.290 --> 00:01:54.090 analyst? Not one that spits out charts, but one 00:01:54.090 --> 00:01:56.269 that asks better questions. Hey everyone, I'm 00:01:56.269 --> 00:01:59.250 Mukundan. And this is Data and AI with Mukundan. 00:01:59.790 --> 00:02:02.810 This show is about learning AI by building it. 00:02:03.349 --> 00:02:06.829 No jargon, no fluff, just honest stories, practical 00:02:06.829 --> 00:02:09.789 tools, and a bit of curiosity along the way. 00:02:11.340 --> 00:02:14.780 Today's story is special. It's the story of a 00:02:14.780 --> 00:02:19.180 tool that started as a side project and turned 00:02:19.180 --> 00:02:22.000 into something that taught me more about thinking 00:02:22.000 --> 00:02:25.479 than any analytics course could ever do that. 00:02:25.819 --> 00:02:29.080 And it's the story of the AI that thinks like 00:02:29.080 --> 00:02:35.099 an analyst and how I, by building it, changed 00:02:35.099 --> 00:02:39.719 how I see curiosity, data, and even... creativity 00:02:39.719 --> 00:02:43.539 itself. Let's look at the problem behind the 00:02:43.539 --> 00:02:46.620 problem. Here's what most people don't talk about. 00:02:47.120 --> 00:02:50.740 The hardest part about data analysis isn't SQL, 00:02:51.020 --> 00:02:55.699 Python, or data visualization. It's asking the 00:02:55.699 --> 00:02:58.879 right question. You can have all the best tools 00:02:58.879 --> 00:03:01.379 in the world, the Tableaus and the Looker, Power 00:03:01.379 --> 00:03:05.060 BI, whatever, right? But if you're solving the 00:03:05.060 --> 00:03:08.849 wrong problem, you're just wasting time. And 00:03:08.849 --> 00:03:13.050 I found myself doing that quite a bit. I'd jump 00:03:13.050 --> 00:03:16.530 into cleaning, filtering, building charts. And 00:03:16.530 --> 00:03:19.250 only halfway through, I'd realize that my question 00:03:19.250 --> 00:03:23.830 wasn't the one that mattered. That realization 00:03:23.830 --> 00:03:28.969 sent me down a big rabbit hole, right? And I 00:03:28.969 --> 00:03:33.129 started reading about exploratory data analysis. 00:03:33.789 --> 00:03:36.310 Asking friends in data science how they would 00:03:36.310 --> 00:03:39.509 approach it. The best analysts I knew weren't 00:03:39.509 --> 00:03:42.949 the fastest coders. They were the ones who would 00:03:42.949 --> 00:03:46.629 pause before touching the data. They'd say things 00:03:46.629 --> 00:03:49.870 like, let's define what success means first. 00:03:50.430 --> 00:03:53.150 Or what decision are we actually trying to make 00:03:53.150 --> 00:03:57.509 here? And that was it. That's what I wanted to 00:03:57.509 --> 00:04:02.189 teach my AI as well. Not to answer, but to ask. 00:04:03.439 --> 00:04:07.639 So I did what, you know, I would, I think most 00:04:07.639 --> 00:04:11.659 people in data would do. I opened my laptop and 00:04:11.659 --> 00:04:14.060 I mean, my most people, I mean, the people who 00:04:14.060 --> 00:04:18.019 would know how to use Python and more specifically 00:04:18.019 --> 00:04:22.740 Streamlit, which is like a front end streaming, 00:04:22.920 --> 00:04:26.180 front end web based library, which would help 00:04:26.180 --> 00:04:32.879 you to create these amazing. and short web applications, 00:04:33.040 --> 00:04:35.500 which is very easy to make. So I opened my laptop 00:04:35.500 --> 00:04:41.720 in this instance, started my streamlet, and wrote 00:04:41.720 --> 00:04:44.100 that first line of code. So when you build something 00:04:44.100 --> 00:04:47.319 like this, it's tempting to focus on the shiny 00:04:47.319 --> 00:04:52.680 part. The UI, the charts, the GPT output. But 00:04:52.680 --> 00:04:57.100 I wanted to focus on the experience. The first 00:04:57.100 --> 00:05:01.439 prototype that I had here was rough. You'd upload 00:05:01.439 --> 00:05:05.279 a CSV file and the app would use GPT -4 to scan 00:05:05.279 --> 00:05:08.139 the column names and generate about 10 questions 00:05:08.139 --> 00:05:12.279 about the data. At first, it was basic. What's 00:05:12.279 --> 00:05:14.699 the average? What's the trend? Nothing exciting. 00:05:15.079 --> 00:05:18.899 But then I started adding prompts that would 00:05:18.899 --> 00:05:24.459 mirror a human curiosity. Things like, which 00:05:24.459 --> 00:05:28.439 variable seems most unpredictable? If this were 00:05:28.439 --> 00:05:31.560 your business, what metric would you track weekly? 00:05:32.399 --> 00:05:37.740 Where might hidden seasonality exist? And suddenly 00:05:37.740 --> 00:05:41.139 the tone shifted. It didn't feel like a chatbot 00:05:41.139 --> 00:05:45.019 anymore. It felt like a mentor. The kind of curious 00:05:45.019 --> 00:05:48.379 teammate who nudges you towards better thinking. 00:05:49.160 --> 00:05:53.339 And one person who actually tested this said 00:05:53.339 --> 00:05:55.360 that it's like the app that is teaching me how 00:05:55.360 --> 00:05:59.410 to explore, not just analyze. That's when I knew 00:05:59.410 --> 00:06:01.310 that I was onto something. Let's look at that. 00:06:01.430 --> 00:06:05.389 You know, technically speaking, here's what happens 00:06:05.389 --> 00:06:09.069 really under the hood. The app reads the metadata, 00:06:09.470 --> 00:06:14.089 column names, types, maybe a small preview of 00:06:14.089 --> 00:06:18.329 values. Now, instead of sending the entire data 00:06:18.329 --> 00:06:22.829 set, which could expose the PII, it summarizes 00:06:22.829 --> 00:06:25.610 the schema, converts it into a structured context. 00:06:26.189 --> 00:06:30.310 and passes that to GPT -4. Then GPT crafts 10 00:06:30.310 --> 00:06:34.089 thoughtful domain -agnostic questions, tuned 00:06:34.089 --> 00:06:39.170 for exploration. Not answers, just sparks. Later 00:06:39.170 --> 00:06:43.250 versions added an objective field, a little text 00:06:43.250 --> 00:06:45.829 box where you could say, I'm exploring customer 00:06:45.829 --> 00:06:49.930 churn, or I want to find high -value users. The 00:06:49.930 --> 00:06:52.930 AI would then adapt its curiosity to match your 00:06:52.930 --> 00:06:58.149 intent. That was version 2. But by version 4, 00:06:58.310 --> 00:07:01.649 I was obsessed with more security. I tend to 00:07:01.649 --> 00:07:04.649 want people uploading data at all. So I built 00:07:04.649 --> 00:07:07.509 a no -upload version as well, wherein you paste 00:07:07.509 --> 00:07:10.930 the schema or sample rows locally. That converts 00:07:10.930 --> 00:07:14.189 it into embeddings and sends only abstracted 00:07:14.189 --> 00:07:18.069 info. Here's what, you're not having any PII 00:07:18.069 --> 00:07:21.029 at this point. No data leaks, no corporate nightmares. 00:07:21.709 --> 00:07:24.939 Because here's what can happen, right? You are 00:07:24.939 --> 00:07:29.959 feeding the AI some kind of sensitive information 00:07:29.959 --> 00:07:32.959 these days. And you may not realize that it may 00:07:32.959 --> 00:07:36.879 be more of a negligence case. You know, you're 00:07:36.879 --> 00:07:39.259 trying to paste something in ChartGPT or Cloud 00:07:39.259 --> 00:07:41.920 or whatever. And this is not an internal company 00:07:41.920 --> 00:07:44.699 tool. This could be something that you just open 00:07:44.699 --> 00:07:47.540 your regular ChartGPT and type in over there. 00:07:48.339 --> 00:07:50.879 But what you may not realize is sometimes you're 00:07:50.879 --> 00:07:54.790 training their models. And by providing the data, 00:07:54.889 --> 00:07:56.529 like let's just say you're working for a company 00:07:56.529 --> 00:07:58.850 and the company name appears over there in the 00:07:58.850 --> 00:08:01.209 data set. Yeah, you're probably trading the data 00:08:01.209 --> 00:08:05.529 for sure. And when you're doing that, it can 00:08:05.529 --> 00:08:08.189 be a problem for the company. The company would 00:08:08.189 --> 00:08:10.810 be looking at it very differently. Maybe they 00:08:10.810 --> 00:08:14.949 find you doing that and there could be consequences, 00:08:15.069 --> 00:08:19.149 which I don't think, I don't know if most companies 00:08:19.149 --> 00:08:22.170 have gotten to that stage yet, but. I feel like 00:08:22.170 --> 00:08:29.290 it might at some point. So the whole point is 00:08:29.290 --> 00:08:33.450 to avoid that kind of scenario. So you're just 00:08:33.450 --> 00:08:36.730 sending out a few samples and remove those PII's. 00:08:37.409 --> 00:08:41.029 And this is why it hit me that this isn't just 00:08:41.029 --> 00:08:44.710 like a coding challenge. It's about designing 00:08:44.710 --> 00:08:49.710 trust. Now, let's look at when AI became a mirror 00:08:49.710 --> 00:08:53.659 to me. Something strange started happening as 00:08:53.659 --> 00:08:57.000 I started testing it. Every time the AI asked 00:08:57.000 --> 00:09:00.460 a question, I felt this urge to defend my assumptions. 00:09:01.200 --> 00:09:03.740 It would say something like, why do you measure 00:09:03.740 --> 00:09:07.559 engagement that way? And I'd catch myself thinking, 00:09:07.700 --> 00:09:10.700 because that's how we've always done it. And 00:09:10.700 --> 00:09:14.980 that right there was the point. The app was making 00:09:14.980 --> 00:09:18.480 me reflect on my own thinking. It wasn't telling 00:09:18.480 --> 00:09:22.639 me that I was wrong. It was making me curious 00:09:22.639 --> 00:09:27.259 again. You know how as kids we ask why about 00:09:27.259 --> 00:09:30.240 everything? Why is the sky blue? Why does 2 plus 00:09:30.240 --> 00:09:34.559 2 make 4? Why can't I eat dessert first? Somewhere 00:09:34.559 --> 00:09:39.340 in adulthood, we lose that instinct. We replace 00:09:39.340 --> 00:09:43.600 curiosity with correctness. And this project 00:09:43.600 --> 00:09:49.809 reawakened that childlike curiosity in me. In 00:09:49.809 --> 00:09:52.370 a world of dashboards and KPIs, that's quite 00:09:52.370 --> 00:09:55.049 rare. The biggest lesson this AI taught me wasn't 00:09:55.049 --> 00:09:59.490 technical. It was philosophical. Curiosity isn't 00:09:59.490 --> 00:10:02.009 something you have. It's something you practice. 00:10:02.649 --> 00:10:05.850 Every time you face a new problem, curiosity 00:10:05.850 --> 00:10:09.190 is the courage to say, I don't know that yet. 00:10:09.490 --> 00:10:14.269 Let's find out. Now, in data analysis, this means 00:10:14.269 --> 00:10:18.559 resisting the urge to rush to a conclusion. It 00:10:18.559 --> 00:10:21.600 means sitting with uncertainty long enough to 00:10:21.600 --> 00:10:23.940 understand what the data is actually trying to 00:10:23.940 --> 00:10:27.460 tell you. And I started applying this outside 00:10:27.460 --> 00:10:31.799 of work too. Instead of reacting, I started asking 00:10:31.799 --> 00:10:35.220 questions in daily life. When something frustrated 00:10:35.220 --> 00:10:40.840 me, I'd ask, what else could this mean? Now, 00:10:40.860 --> 00:10:44.000 when I got feedback, I'd ask, what can I learn 00:10:44.000 --> 00:10:49.370 from this? That same mindset. Help me make my 00:10:49.370 --> 00:10:54.110 AI better. We often talk about teaching AI to 00:10:54.110 --> 00:10:57.370 think like humans. But what if we flipped it? 00:10:57.669 --> 00:11:02.190 What if the goal is to help humans think a little 00:11:02.190 --> 00:11:07.669 more like AI, patiently, systematically, with 00:11:07.669 --> 00:11:13.549 curiosity, but also humility? The line between 00:11:13.549 --> 00:11:16.919 human and machine curiosity is... thinner than 00:11:16.919 --> 00:11:20.679 it seems. Now, AI asks questions without ego. 00:11:21.600 --> 00:11:26.159 Humans sometimes, well, they don't. And maybe 00:11:26.159 --> 00:11:29.299 that's the balance we need in the age of intelligent 00:11:29.299 --> 00:11:33.220 systems. Not machines that replace us, but ones 00:11:33.220 --> 00:11:37.480 that remind us to stay curious. Now, when I built 00:11:37.480 --> 00:11:40.139 this project, I thought it was a data tool. But 00:11:40.139 --> 00:11:43.889 looking back, it was a reflection tool. It showed 00:11:43.889 --> 00:11:46.549 me that asking better questions isn't just the 00:11:46.549 --> 00:11:49.850 key to better data. It's the key to better decisions, 00:11:50.090 --> 00:11:52.789 better creativity, and maybe even better living. 00:11:52.950 --> 00:11:54.870 If there's one thing I want you to take away 00:11:54.870 --> 00:11:58.330 from this episode, it's this. Don't just build 00:11:58.330 --> 00:12:02.289 tools that save time. Build tools that make you 00:12:02.289 --> 00:12:07.269 curious again. Because curiosity is the hidden 00:12:07.269 --> 00:12:11.029 engine of every great analysis, every insight. 00:12:11.759 --> 00:12:15.720 and every invention. And if you ever feel stuck 00:12:15.720 --> 00:12:19.460 with data, with work, or with life, maybe what 00:12:19.460 --> 00:12:23.399 you need isn't another answer. Maybe what you 00:12:23.399 --> 00:12:27.759 need is a better question. Well, thanks for listening. 00:12:27.919 --> 00:12:30.519 You can try the AI app yourself. It's linked 00:12:30.519 --> 00:12:34.700 in the show notes. And if you linked... Well, 00:12:34.799 --> 00:12:37.179 I mean, if you like this show, if you like this 00:12:37.179 --> 00:12:39.960 story... share it with someone who's forgotten 00:12:39.960 --> 00:12:42.179 what curiosity feels like before you go i want 00:12:42.179 --> 00:12:44.320 to take this time to thank my affiliate partners 00:12:44.320 --> 00:12:48.019 riverside fm which i use for recording my podcast 00:12:48.019 --> 00:12:51.799 with its amazing ai editing features where i 00:12:51.799 --> 00:12:55.080 can edit my podcast and remove any kind of cider 00:12:55.080 --> 00:12:58.440 .ai for any research i wanted to do with the 00:12:58.440 --> 00:13:02.019 topics i want to talk about and of course my 00:13:02.019 --> 00:13:06.639 podcast hosting platform rss where i host my 00:13:06.639 --> 00:13:09.340 podcast and distribute it to multiple channels 00:13:09.340 --> 00:13:13.100 and using their dynamic ads feature anybody can 00:13:13.100 --> 00:13:15.620 get paid if you just have 10 downloads a month 00:13:15.620 --> 00:13:17.919 so just 10 downloads a month is all you need 00:13:17.919 --> 00:13:23.759 to start making money and well that's it from 00:13:23.759 --> 00:13:26.379 this episode i'm mukundan and this is data and 00:13:26.379 --> 00:13:28.980 ai with mukundan see you next hey it's mukundan 00:13:28.980 --> 00:13:31.679 if this episode helped you two tiny favors that 00:13:31.679 --> 00:13:34.360 make a huge difference rate the show five stars 00:13:34.990 --> 00:13:37.769 On Spotify, you can just open the show page and 00:13:37.769 --> 00:13:40.350 tap the star button. On Apple Podcasts, scroll 00:13:40.350 --> 00:13:42.690 to the bottom of the show page and tap 5 stars. 00:13:42.889 --> 00:13:45.850 Also, leave a one -to -line review on Apple Podcasts. 00:13:46.769 --> 00:13:49.309 Tell me one takeaway. I read every single one 00:13:49.309 --> 00:13:51.009 and it helps more people find the show.