WEBVTT 00:00:00.000 --> 00:00:02.620 Two months ago I shipped a tiny AI tool that 00:00:02.620 --> 00:00:05.339 looks at a CSV file and asks better questions 00:00:05.339 --> 00:00:10.359 than I can on day two at a new job. It went way 00:00:10.359 --> 00:00:12.779 further than I had expected and it showed me 00:00:12.779 --> 00:00:17.440 what I had missed. So I rebuilt it. Today I'll 00:00:17.440 --> 00:00:19.920 share what worked, what didn't and I'll give 00:00:19.920 --> 00:00:23.940 you a short quiz, a discussion prompt and a checklist 00:00:23.940 --> 00:00:33.509 that you can use. Hey, I'm Mukundan. This show 00:00:33.509 --> 00:00:35.810 is about solving real problems with small, useful 00:00:35.810 --> 00:00:39.810 AI. Things that you can actually use today. Each 00:00:39.810 --> 00:00:43.490 week, we pick one problem, build a simple workflow 00:00:43.490 --> 00:00:46.130 or tool, and talk through the decisions about 00:00:46.130 --> 00:00:49.250 what to automate, how to check quality, and how 00:00:49.250 --> 00:00:52.710 to make it reliable. If you're a builder, creator, 00:00:52.929 --> 00:00:56.329 or just AI curious, you'll leave with the steps 00:00:56.329 --> 00:01:00.070 that you can copy tonight. Hit follow and start 00:01:00.070 --> 00:01:02.460 with the Startup Pack episodes. in the show notes. 00:01:02.579 --> 00:01:04.299 Hey everyone, welcome back to the Data and AI 00:01:04.299 --> 00:01:08.840 with Mukundan show. Today, we'll talk about how 00:01:08.840 --> 00:01:14.500 I built an AI tool that helped look at CSV files 00:01:14.500 --> 00:01:18.420 and Excel files and generate questions from it 00:01:18.420 --> 00:01:21.819 and how I took it to the next level. This is 00:01:21.819 --> 00:01:25.680 about that story. You join a new team. Here's 00:01:25.680 --> 00:01:29.620 why this exists. I call it the day two problem. 00:01:30.890 --> 00:01:35.209 You join a new team in a company and maybe let's 00:01:35.209 --> 00:01:38.609 just say, or even you join a new company and 00:01:38.609 --> 00:01:41.650 someone will send you this data set without any 00:01:41.650 --> 00:01:45.810 context and no goals about what you want to achieve 00:01:45.810 --> 00:01:49.170 with this data. It's just, just tell you with 00:01:49.170 --> 00:01:50.790 pressure, right? Just pressure. That's all it 00:01:50.790 --> 00:01:57.370 is. And version one of my tool, when I first 00:01:57.370 --> 00:02:01.200 did this, you know, this project before. It helped 00:02:01.200 --> 00:02:05.799 me by generating 10 smart exploratory data analysis 00:02:05.799 --> 00:02:09.180 questions like the EDA questions, each with a 00:02:09.180 --> 00:02:12.840 short why this matters. And it gave me like a 00:02:12.840 --> 00:02:17.740 plan instead of me panicking about it, right? 00:02:18.620 --> 00:02:21.639 So I built this tool before. I called this version 00:02:21.639 --> 00:02:28.099 one, right? And what happened was I... created 00:02:28.099 --> 00:02:31.060 this like exploratory data analysis tool what 00:02:31.060 --> 00:02:34.800 basically it would take you take your csv file 00:02:34.800 --> 00:02:39.099 and then it would generate questions from it 00:02:39.099 --> 00:02:43.620 give you like 10 smart questions to ask ask this 00:02:43.620 --> 00:02:47.300 data set right that will help you really to understand 00:02:47.300 --> 00:02:50.300 the data and it will also give you like a why 00:02:50.300 --> 00:02:54.120 this matters why this question matters statement 00:02:54.120 --> 00:02:57.960 as well so it gives you like a plan right so 00:02:57.960 --> 00:03:00.219 to speak like it give you like what what to do 00:03:00.219 --> 00:03:02.180 gives you a plan instead of like you know you're 00:03:02.180 --> 00:03:08.460 panicking about it um so when i did this when 00:03:08.460 --> 00:03:10.840 i did the first version of this it went viral 00:03:10.840 --> 00:03:14.580 a lot of people were able to resonate with it 00:03:14.580 --> 00:03:18.879 uh some even you know some reacted to the blog 00:03:18.879 --> 00:03:21.479 post i'd written it was a blog post with code 00:03:21.479 --> 00:03:25.479 and i'll link to the blog post as well So I'll 00:03:25.479 --> 00:03:29.060 tell you what worked, why it went viral and what 00:03:29.060 --> 00:03:33.599 was missing in this version one. So people liked 00:03:33.599 --> 00:03:37.479 three things. They liked that there was a fast 00:03:37.479 --> 00:03:42.719 grip, right? I mean, you would be able to get 00:03:42.719 --> 00:03:46.800 really useful questions from the data in under 00:03:46.800 --> 00:03:51.740 a minute. And that's really quick, right? Like, 00:03:51.740 --> 00:03:54.060 I mean, otherwise we're just spending hours and 00:03:54.060 --> 00:03:56.360 hours, maybe days and days, just looking at the 00:03:56.360 --> 00:03:59.699 data set without knowing what to answer. And 00:03:59.699 --> 00:04:02.080 especially when you're looking at it fresh, you 00:04:02.080 --> 00:04:04.300 don't know what's going on and you don't know 00:04:04.300 --> 00:04:07.500 what question to ask. We don't know how to answer 00:04:07.500 --> 00:04:09.219 that because A, you don't know what question 00:04:09.219 --> 00:04:11.259 to ask really, right? You don't want to be doing 00:04:11.259 --> 00:04:14.819 count distinct, right? It's a SQL query if you're 00:04:14.819 --> 00:04:17.220 not familiar with it. Count distinct will basically... 00:04:19.660 --> 00:04:22.759 in a way just count the unique records or something 00:04:22.759 --> 00:04:27.720 right of whatever is the number of whatever is 00:04:27.720 --> 00:04:29.660 the table you're looking at so let's say you 00:04:29.660 --> 00:04:35.579 want to look at the number of you know let's 00:04:35.579 --> 00:04:37.319 say you want to look at the number of cars and 00:04:37.319 --> 00:04:40.720 you would do that by let's say car id right you 00:04:40.720 --> 00:04:43.019 have something called as a car id field so you 00:04:43.019 --> 00:04:46.500 would be using count distinct to see how many 00:04:48.000 --> 00:04:52.279 you know unique cars there are in this showroom 00:04:52.279 --> 00:04:57.199 for example right so that was something that 00:04:57.199 --> 00:05:01.060 will not really help but we want to be asking 00:05:01.060 --> 00:05:05.399 some more useful questions and I'll talk about 00:05:05.399 --> 00:05:07.819 what questions those are and everything as we 00:05:07.819 --> 00:05:10.579 go down but I just wanted to give you some background 00:05:10.579 --> 00:05:15.199 about this so the other thing people liked was 00:05:15.199 --> 00:05:20.670 like it acted like a partner gave you a vibe 00:05:20.670 --> 00:05:26.009 of a partner where it nudges but not doesn't 00:05:26.009 --> 00:05:28.290 doesn't give you like a dashboard dump right 00:05:28.290 --> 00:05:31.470 it's not like dumping everything on you it's 00:05:31.470 --> 00:05:34.189 just like being your partner right you support 00:05:34.189 --> 00:05:37.209 your partner and so it's it's basically giving 00:05:37.209 --> 00:05:39.930 you those questions and then once you have those 00:05:39.930 --> 00:05:43.470 questions all you have to do is just uh you know 00:05:43.470 --> 00:05:46.279 use your analytical skills to answer those questions 00:05:46.279 --> 00:05:50.800 and there was also a respect for time that's 00:05:50.800 --> 00:05:54.019 the third thing people liked because it was short 00:05:54.019 --> 00:05:59.459 clear and it was human so to recap people like 00:05:59.459 --> 00:06:03.800 three things because it was a fast b acted like 00:06:03.800 --> 00:06:07.550 your partner and three respected your time So 00:06:07.550 --> 00:06:10.009 I feel like I still missed three key pieces. 00:06:10.209 --> 00:06:12.990 That's what people said. I mean, that's what 00:06:12.990 --> 00:06:19.709 I believed it as well. I hadn't mentioned what 00:06:19.709 --> 00:06:23.089 was the business objective up front and why am 00:06:23.089 --> 00:06:26.189 I doing this exercise? Why am I doing this little 00:06:26.189 --> 00:06:28.029 project, this AI tool? Why did I build that? 00:06:28.649 --> 00:06:32.110 I didn't really mention any objective up front. 00:06:33.350 --> 00:06:40.040 And I also did not, focus on any columns right 00:06:40.040 --> 00:06:42.699 i treated basically every feature the same and 00:06:42.699 --> 00:06:47.360 everyone knows that if you treat every column 00:06:47.360 --> 00:06:52.480 in your table as the same how is that going to 00:06:52.480 --> 00:06:59.060 you know really help you right um and i also 00:06:59.060 --> 00:07:02.620 wanted to give an option to export the data set 00:07:02.620 --> 00:07:04.839 or export the you know questions and everything 00:07:04.839 --> 00:07:10.050 so that was missing as well so there's no export 00:07:10.050 --> 00:07:12.490 feature to available no export feature which 00:07:12.490 --> 00:07:16.589 was available to drop it into slack or even into 00:07:16.589 --> 00:07:19.949 notion or your jira board right so that was uh 00:07:19.949 --> 00:07:22.610 something i felt was missing so i wanted to take 00:07:22.610 --> 00:07:25.350 that to the next level here were some lessons 00:07:25.350 --> 00:07:30.550 that helped me rebuild this so i wanted to first 00:07:30.550 --> 00:07:35.550 put the decision in the room first spend attention 00:07:35.550 --> 00:07:39.129 like money attention in this case would be time 00:07:39.129 --> 00:07:42.230 and time is money so focus on you know what i 00:07:42.230 --> 00:07:45.750 wanted to do um what problem i wanted to solve 00:07:45.750 --> 00:07:49.910 and like a little one page brief so that the 00:07:49.910 --> 00:07:52.029 work actually moves here's how the rebuild actually 00:07:52.029 --> 00:07:55.149 took place the three simple features now before 00:07:55.149 --> 00:07:58.769 anything else the tool asks what outcome are 00:07:58.769 --> 00:08:00.949 you trying to move so this is the objective input 00:08:00.949 --> 00:08:05.910 the y box right so let's talk about that so just 00:08:05.910 --> 00:08:07.949 to ask yourself like what outcome are you trying 00:08:07.949 --> 00:08:13.209 to move so example is 30 do you want to example 00:08:13.209 --> 00:08:15.230 as you want to reduce 30 day churn for monthly 00:08:15.230 --> 00:08:18.610 plan users by 15 an optional mode would be risk 00:08:18.610 --> 00:08:21.629 opportunity anomaly or quick wins to shape the 00:08:21.629 --> 00:08:25.569 tone result questions became specific and useful 00:08:25.569 --> 00:08:28.670 and then i wanted to also do a column focus the 00:08:28.670 --> 00:08:31.889 second simple feature one was the objective input 00:08:31.889 --> 00:08:35.690 so what outcome i'm trying to move right and 00:08:35.690 --> 00:08:39.690 the second one more on the columns and this is 00:08:39.690 --> 00:08:43.370 specifically how much columns i want five to 00:08:43.370 --> 00:08:45.690 eight columns only is what i chose and up to 00:08:45.690 --> 00:08:49.909 eight columns or except smart hints right ask 00:08:49.909 --> 00:08:55.169 for that so you should want to be able to look 00:08:55.169 --> 00:08:57.450 at the data quickly if you look at 30 columns 00:08:57.450 --> 00:09:00.950 together like you want to pick out every feature 00:09:01.480 --> 00:09:05.039 how is that going to help so that is not a time 00:09:05.039 --> 00:09:09.799 saving tool or maybe it saves a little bit but 00:09:09.799 --> 00:09:12.639 not as much as you would want right so the column 00:09:12.639 --> 00:09:16.259 focus where we picked up to eight columns or 00:09:16.259 --> 00:09:19.179 accept any smart hints and the tool concentrates 00:09:19.179 --> 00:09:23.919 its thinking budget on what matters and warns 00:09:23.919 --> 00:09:27.809 about any leakage or messy labels So that was 00:09:27.809 --> 00:09:30.409 number two. And then number three was markdown 00:09:30.409 --> 00:09:33.669 export. So ship the plan. So one click and you 00:09:33.669 --> 00:09:36.230 get a clean brief. So the objective in your own 00:09:36.230 --> 00:09:39.009 words, you'll mention top 10 questions and why 00:09:39.009 --> 00:09:41.330 this matters. Those three that was already there. 00:09:42.669 --> 00:09:46.230 And data health checks, quick analysis to run 00:09:46.230 --> 00:09:52.750 risks and next decisions. And also paste it into 00:09:52.750 --> 00:09:56.840 a Slack, Notion and Jira and go. right so before 00:09:56.840 --> 00:09:58.759 anything else like i mentioned like these were 00:09:58.759 --> 00:10:02.299 the three rebuild features simple features that 00:10:02.299 --> 00:10:05.399 i've worked on the objective input where we're 00:10:05.399 --> 00:10:10.100 trying to uh where you're trying to figure out 00:10:10.100 --> 00:10:12.460 what outcome you're trying to move the column 00:10:12.460 --> 00:10:14.940 focus where you're picking up to eight columns 00:10:14.940 --> 00:10:18.720 or accepting any smart hints the markdown export 00:10:18.720 --> 00:10:23.700 right where you're able to export the data set 00:10:23.700 --> 00:10:26.500 and export the the questions really like objective 00:10:26.500 --> 00:10:28.980 the 10 questions why this matters the data health 00:10:28.980 --> 00:10:31.840 checks quick analysis to run and risks and any 00:10:31.840 --> 00:10:34.799 next decisions all right so here's here's a fun 00:10:34.799 --> 00:10:37.899 fun session for you lightning quiz that i have 00:10:37.899 --> 00:10:40.100 for you here's a quick quiz that i just prepared 00:10:40.100 --> 00:10:42.860 and i'll read the answers right after i'll give 00:10:42.860 --> 00:10:46.220 you some time to think if you're if you're listening 00:10:46.220 --> 00:10:50.620 um so that like you know um i don't want to directly 00:10:50.620 --> 00:10:52.559 prompt the answers and maybe just give you a 00:10:52.559 --> 00:10:54.259 few seconds to think. First, I'll just give you 00:10:54.259 --> 00:10:56.399 the questions and then you can think about it 00:10:56.399 --> 00:10:59.259 and then, you know, obviously give me the answers. 00:11:00.539 --> 00:11:03.019 Well, I mean, the answers I will be giving you, 00:11:03.080 --> 00:11:07.580 but you can also come up with your own. You can 00:11:07.580 --> 00:11:11.500 also send me on chat. Just so that I know you're 00:11:11.500 --> 00:11:14.840 listening and you're paying attention. So yeah, 00:11:14.899 --> 00:11:16.740 just quick quiz. I'll read the answers right 00:11:16.740 --> 00:11:20.120 after. So pause if you need more time. Question 00:11:20.120 --> 00:11:25.799 one, true or false? Start exploratory data analysis 00:11:25.799 --> 00:11:31.360 or EDA by uploading the data and exploring the 00:11:31.360 --> 00:11:33.720 charts, right? So I'll just repeat this again. 00:11:33.879 --> 00:11:40.379 True or false? Start EDA by uploading data and 00:11:40.379 --> 00:11:43.799 exploring charts. So it's telling you that as 00:11:43.799 --> 00:11:47.429 soon as you upload the data, and explores the 00:11:47.429 --> 00:11:50.629 charts, you need to start EDA. Question 2. You 00:11:50.629 --> 00:11:54.370 have 50 columns. What do you focus on? 5 to 8, 00:11:54.529 --> 00:11:58.710 that's option A. Option B, 20 to 25. Option C, 00:11:58.769 --> 00:12:02.110 all 50. What comes first? Writing a business 00:12:02.110 --> 00:12:07.049 objective, option A. Cleaning every column, option 00:12:07.049 --> 00:12:10.950 B. And option C, running 20 plots. Question 4. 00:12:11.129 --> 00:12:14.629 If you were to ask, if you were asked, to predict 00:12:14.629 --> 00:12:18.269 the 30 -day churn and you're asked what was the 00:12:18.269 --> 00:12:21.169 leakage here would it be 10 -yard days on day 00:12:21.169 --> 00:12:25.190 one would it be churned in 30 days or would it 00:12:25.190 --> 00:12:28.509 be region so just to like reiterate leakage means 00:12:28.509 --> 00:12:31.750 really the target variable right if you're asked 00:12:31.750 --> 00:12:34.590 to be predicting the 30 -day churn what would 00:12:34.590 --> 00:12:36.529 be the target variable would it be the 10 -yard 00:12:36.529 --> 00:12:40.490 days on day one would it be churned in 30 days 00:12:40.490 --> 00:12:45.320 or would it be So there's three options, right? 00:12:45.659 --> 00:12:47.720 I'll just repeat the options again just so that 00:12:47.720 --> 00:12:51.960 you guys are clear. Option A, 10 -year days on 00:12:51.960 --> 00:12:55.019 day one. Option B, churned in 30 days. And option 00:12:55.019 --> 00:12:57.879 C, region. And finally, question five. So just 00:12:57.879 --> 00:13:02.240 to reiterate, what is the best artifact after 00:13:02.240 --> 00:13:05.700 an exploratory data analysis? Is it option A, 00:13:05.840 --> 00:13:10.100 dashboard gallery? Is it option B, a one -page 00:13:10.100 --> 00:13:14.159 markdown brief? Or is it option C, a raw notebook? 00:13:14.580 --> 00:13:19.799 Question 1, true or false? Start ADA. Would you 00:13:19.799 --> 00:13:22.840 start ADA by uploading the data and exploring 00:13:22.840 --> 00:13:27.620 charts? Question 2, when you have 50 columns, 00:13:27.899 --> 00:13:31.700 what would you focus on? 5 to 8 columns, option 00:13:31.700 --> 00:13:35.120 A. Option B, 20 to 25 columns. Or option C, all 00:13:35.120 --> 00:13:40.860 50 columns. Question 3, what comes first? A, 00:13:40.879 --> 00:13:45.220 writing a business objective. B, cleaning every 00:13:45.220 --> 00:13:49.759 column. Or C, running 20 plots. And question 00:13:49.759 --> 00:13:53.080 five, what would be the best artifact to give 00:13:53.080 --> 00:13:57.500 after an EDA? Would it be a dashboard gallery? 00:13:57.740 --> 00:14:01.379 Would it be a one -page markdown brief? As option 00:14:01.379 --> 00:14:04.039 B is one -page markdown and option A is dashboard 00:14:04.039 --> 00:14:07.840 gallery. And option C, would it be a raw notebook? 00:14:08.570 --> 00:14:11.889 here we go with the answers so the answer one 00:14:11.889 --> 00:14:16.950 to true or false it's false so you would first 00:14:16.950 --> 00:14:20.850 start with the decision right you wouldn't directly 00:14:20.850 --> 00:14:23.070 just do eda that's what it means really upload 00:14:23.070 --> 00:14:25.909 data and to ed that doesn't it's not how it works 00:14:25.909 --> 00:14:28.850 you need to be understanding the business question 00:14:28.850 --> 00:14:31.950 and everything so don't directly jump into exploring 00:14:31.950 --> 00:14:36.970 charts um answer two is five to eight columns 00:14:38.570 --> 00:14:41.289 so i asked what what does the you know how many 00:14:41.289 --> 00:14:43.450 columns would you choose like if you have 50 00:14:43.450 --> 00:14:46.330 columns which would you focus on the answer is 00:14:46.330 --> 00:14:49.490 five to eight columns you don't want to be overburdening 00:14:49.490 --> 00:14:51.929 yourself right so that's something you would 00:14:51.929 --> 00:14:56.090 tell the ai to do as well answer three what comes 00:14:56.090 --> 00:14:59.210 first between writing a business objective cleaning 00:14:59.210 --> 00:15:02.669 every column or running 20 plots the objective 00:15:02.669 --> 00:15:06.389 comes first so the options were writing a business 00:15:06.389 --> 00:15:09.139 objective cleaning every column or running 20 00:15:09.139 --> 00:15:12.840 plots. So objective, writing a business objective 00:15:12.840 --> 00:15:15.480 comes first. Question four was predicting 30 00:15:15.480 --> 00:15:18.200 -day churn. What was the output variable? Was 00:15:18.200 --> 00:15:20.960 it 10 -year days on day one? Was it churned in 00:15:20.960 --> 00:15:24.340 30 days or was it region? The answer is churned 00:15:24.340 --> 00:15:25.960 in 30 days because that is the target variable. 00:15:26.279 --> 00:15:29.919 Question five, what was the best artifact after 00:15:29.919 --> 00:15:34.720 an exploratory data analysis? Is it a dashboard 00:15:34.720 --> 00:15:37.370 gallery? one page markdown brief or raw notebook 00:15:37.370 --> 00:15:40.990 answer is one page brief so those were the answers 00:15:40.990 --> 00:15:45.870 so again just to reiterate answer one was false 00:15:45.870 --> 00:15:48.450 so you don't directly start eda you first start 00:15:48.450 --> 00:15:51.730 with the the business question really what you 00:15:51.730 --> 00:15:55.529 decided to do so that is one answer two was if 00:15:55.529 --> 00:15:57.730 you have 50 columns you focus on five to eight 00:15:57.730 --> 00:16:02.230 columns first so question three what comes first 00:16:03.889 --> 00:16:06.289 between writing a business objective, cleaning 00:16:06.289 --> 00:16:08.769 every column and running 20 plots, the objective 00:16:08.769 --> 00:16:12.269 comes first. And answer four, when you're trying 00:16:12.269 --> 00:16:14.710 to predict the 30 -day churn, what is the output 00:16:14.710 --> 00:16:19.009 variable? Is it tenure days on day one, churned 00:16:19.009 --> 00:16:21.210 in 30 days or region? The answer is churned in 00:16:21.210 --> 00:16:24.889 30 days. And answer five, best artifact after 00:16:24.889 --> 00:16:28.700 EDA? A dashboard, gallery, one -page markdown 00:16:28.700 --> 00:16:30.980 brief, or raw notebook? The answer is a one -page 00:16:30.980 --> 00:16:34.840 brief is the best artifact after an EDA. So that 00:16:34.840 --> 00:16:38.419 was the quiz. If you scored a 5 on 5, you're 00:16:38.419 --> 00:16:41.559 ready. You're ready to be the expert data analyst 00:16:41.559 --> 00:16:46.419 yourself. If you scored between 3, 2, 4, and 00:16:46.419 --> 00:16:51.519 4 or 5, just use the checklist next. And if you 00:16:51.519 --> 00:16:55.980 scored between 0 to 2 out of 5, three to four 00:16:55.980 --> 00:16:57.620 if you scored then just use the checklist and 00:16:57.620 --> 00:17:03.580 zero to two if you scored um don't worry this 00:17:03.580 --> 00:17:05.980 checklist will help a lot all right so discussion 00:17:05.980 --> 00:17:09.759 time i want to just like do a little discussion 00:17:09.759 --> 00:17:12.420 with you and this is this is a discussion that 00:17:12.420 --> 00:17:15.099 will be held in my sub stack so basically sub 00:17:15.099 --> 00:17:17.339 stack is uh if you're not familiar with sub stack 00:17:17.339 --> 00:17:21.619 is it's one of the great content creating websites 00:17:21.619 --> 00:17:27.269 um it gives content creators a platform to be 00:17:27.269 --> 00:17:33.109 heard uh through primarily i would say their 00:17:33.109 --> 00:17:36.210 written content yeah sure you get like a lot 00:17:36.210 --> 00:17:40.430 of audio and video uh podcast hosts or whatever 00:17:40.430 --> 00:17:44.349 but i think primarily what i've seen what works 00:17:44.349 --> 00:17:47.650 a lot on substack is written text so that's a 00:17:47.650 --> 00:17:50.849 good platform for that a lot of creators uh are 00:17:50.849 --> 00:17:53.190 getting their voices heard on substack but anyway 00:17:54.539 --> 00:17:58.500 I wanted to also invite a discussion on Substack. 00:17:59.079 --> 00:18:02.279 So here's what I want to discuss about. I want 00:18:02.279 --> 00:18:07.660 to hear your story and your process. Let's say 00:18:07.660 --> 00:18:11.279 you joined that new company, your new team on 00:18:11.279 --> 00:18:14.200 day two or day three, maybe week one for that 00:18:14.200 --> 00:18:18.180 matter. You were asked to look at dataset. And, 00:18:18.200 --> 00:18:21.720 you know, they just asked you like, you know, 00:18:21.740 --> 00:18:25.609 tell me. this is a messy data what will you come 00:18:25.609 --> 00:18:30.410 up with that is a lot of stress um because you're 00:18:30.410 --> 00:18:32.849 trying to make a new you know an an impression 00:18:32.849 --> 00:18:36.589 on this team that you want to be like on top 00:18:36.589 --> 00:18:38.470 right like you want to make a great impression 00:18:38.470 --> 00:18:42.450 you want to feel that they made a good choice 00:18:42.450 --> 00:18:45.890 in hiring you uh for their team so that's why 00:18:45.890 --> 00:18:49.869 i say day two data set horror story so yeah just 00:18:50.509 --> 00:18:53.009 I want to hear your story in this. How do you 00:18:53.009 --> 00:18:57.170 handle a situation like this? The other thing 00:18:57.170 --> 00:19:00.369 I wanted to share, I wanted to discuss is what 00:19:00.369 --> 00:19:02.930 is one metric that you would move in 30 days, 00:19:03.089 --> 00:19:09.269 in your first 30 days? And why? Which metric 00:19:09.269 --> 00:19:13.029 you would move and why? And the third thing, 00:19:13.170 --> 00:19:15.289 the five to eight columns that you would start 00:19:15.289 --> 00:19:17.450 with right now. So let's just say you're looking 00:19:17.450 --> 00:19:22.869 at maybe... a churn data set what would be five 00:19:22.869 --> 00:19:24.710 to eight columns that you'd want to look at right 00:19:24.710 --> 00:19:28.109 now anything what is your day two data set horror 00:19:28.109 --> 00:19:31.190 story what is one metric that you would move 00:19:31.190 --> 00:19:35.630 in your first 30 days and also in 30 days in 00:19:35.630 --> 00:19:39.170 general and why would you move that and third 00:19:39.170 --> 00:19:42.130 what are the five to eight columns that you would 00:19:42.130 --> 00:19:46.650 start with right now and an example here Could 00:19:46.650 --> 00:19:48.650 be anything that you would want to talk about 00:19:48.650 --> 00:19:51.150 your organization or anything that you, you know, 00:19:51.150 --> 00:19:52.970 are passionate about with this five to eight 00:19:52.970 --> 00:19:56.890 columns. I would, you know, definitely appreciate 00:19:56.890 --> 00:20:00.190 like if you would also add a comment in this 00:20:00.190 --> 00:20:04.589 discussion with an okay to read, like a comment 00:20:04.589 --> 00:20:08.549 in quotes saying okay to read. If I can feature 00:20:08.549 --> 00:20:11.809 your reply the next episode or even anonymous 00:20:11.809 --> 00:20:16.599 if you prefer, right? I read every comment. So 00:20:16.599 --> 00:20:18.779 join the discussion and I'll link the discussion 00:20:18.779 --> 00:20:21.799 in the show notes. So that was a discussion that 00:20:21.799 --> 00:20:24.539 I wanted to have. Now I wanted to give a quick 00:20:24.539 --> 00:20:27.059 walkthrough of how an app would work like this. 00:20:27.440 --> 00:20:29.559 So let's just say that you're talking about a 00:20:29.559 --> 00:20:33.319 subscription app. That's the scenario. The objective 00:20:33.319 --> 00:20:37.400 type is reduce the 30 -day churn for monthly 00:20:37.400 --> 00:20:43.059 planned users by 15%. The focus columns, the 00:20:43.059 --> 00:20:44.740 columns you want to focus on would be planned 00:20:44.740 --> 00:20:49.130 type. a signup channel, 10 -year days, region, 00:20:49.450 --> 00:20:55.910 device, discount applied, and support tickets, 00:20:56.049 --> 00:20:58.289 which is number of support tickets and first 00:20:58.289 --> 00:21:02.609 week activity rate. So here are some of the questions 00:21:02.609 --> 00:21:05.509 the tool generates with and why. But let me just 00:21:05.509 --> 00:21:08.369 go over the columns again. And it may be the 00:21:08.369 --> 00:21:10.109 scenario. So the scenario is let's say you want 00:21:10.109 --> 00:21:12.529 to build a subscription app, right? And this 00:21:12.529 --> 00:21:16.579 could be... let's say you're working for a subscription 00:21:16.579 --> 00:21:21.220 -based company and their primary revenue is coming 00:21:21.220 --> 00:21:23.900 from subscriptions. For example, you can take 00:21:23.900 --> 00:21:25.519 your Netflix or whatever, they're subscription 00:21:25.519 --> 00:21:30.619 -based, and your objective for them is you want 00:21:30.619 --> 00:21:34.119 to reduce 30 -day churn for them for monthly 00:21:34.119 --> 00:21:38.559 plan users by 15%, right? So that would be an 00:21:38.559 --> 00:21:42.960 objective. the columns you want to focus on is 00:21:42.960 --> 00:21:45.740 plan type sign up channel how did they sign up 00:21:45.740 --> 00:21:47.640 on really like what is the plan type is it a 00:21:47.640 --> 00:21:50.779 basic plan or whatever tenure days how many days 00:21:50.779 --> 00:21:54.579 they've been a subscriber um which region where 00:21:54.579 --> 00:21:57.099 in the world are they from which device are they 00:21:57.099 --> 00:22:00.119 sign up on or which device are they using so 00:22:00.119 --> 00:22:06.400 to to watch the shows in this case um and if 00:22:06.400 --> 00:22:10.039 they had any discount applied so where i in the 00:22:10.039 --> 00:22:12.559 united states at least i don't think netflix 00:22:12.559 --> 00:22:15.700 has any discount anymore um but that being said 00:22:15.700 --> 00:22:17.380 like you know it could be there anywhere else 00:22:17.380 --> 00:22:20.519 in the world um so discount applied is another 00:22:20.519 --> 00:22:23.720 column here and support tickets is number of 00:22:23.720 --> 00:22:29.480 support tickets um that may be this this user 00:22:29.480 --> 00:22:33.500 has created this user has requested right or 00:22:33.500 --> 00:22:36.349 in the first week activity rate like They started 00:22:36.349 --> 00:22:41.450 using it and how much time they've been watching 00:22:41.450 --> 00:22:46.970 it in the first week. So this is, I'm just giving 00:22:46.970 --> 00:22:49.710 Netflix as an example, but whatever, right? Could 00:22:49.710 --> 00:22:51.990 be any kind of subscription app. The objective 00:22:51.990 --> 00:22:55.109 really is to reduce the 30 -day churn for monthly 00:22:55.109 --> 00:22:59.950 plan users by about 15%. So now let's look at 00:22:59.950 --> 00:23:02.049 some of the questions that the tool has generated 00:23:02.049 --> 00:23:05.069 with the why. The questions it came up with was, 00:23:05.960 --> 00:23:09.559 are churn rates higher for paid social versus 00:23:09.559 --> 00:23:12.940 organic search? And the why is because it's a 00:23:12.940 --> 00:23:17.359 customer acquisition cost or a CAC payback risk, 00:23:17.440 --> 00:23:21.140 right? Because organic search, you're not really 00:23:21.140 --> 00:23:23.519 losing any money. Paid social, you're paying 00:23:23.519 --> 00:23:27.799 for that, the visibility. So if a consumer who's 00:23:27.799 --> 00:23:30.059 joined in through organic search and you're losing 00:23:30.059 --> 00:23:33.819 that customer, quick versus you're losing that 00:23:33.819 --> 00:23:36.880 customer who joined in through paid social so 00:23:36.880 --> 00:23:39.599 that's a problem like i mean if you're losing 00:23:39.599 --> 00:23:41.519 more on paid social versus on organic search 00:23:41.519 --> 00:23:45.819 then um your paid social strategy is clearly 00:23:45.819 --> 00:23:48.079 not working so that's going to be a big big uh 00:23:48.079 --> 00:23:51.240 problem on your customer acquisition cost will 00:23:51.240 --> 00:23:55.119 just rise higher right um no i mean customer 00:23:55.119 --> 00:23:56.900 acquisition cost is one thing but i think what 00:23:56.900 --> 00:24:00.140 what you're gonna face especially here in this 00:24:00.140 --> 00:24:03.240 case that i'm thinking about it is well you spend 00:24:03.240 --> 00:24:07.200 this money now you want to make money from this 00:24:07.200 --> 00:24:11.759 customer so like a payback so to speak so that 00:24:11.759 --> 00:24:13.920 would affect the payback that you get from the 00:24:13.920 --> 00:24:18.660 customer because well their return is low that 00:24:18.660 --> 00:24:22.059 is one question the second question is do sign 00:24:22.059 --> 00:24:25.359 up discounts lead to lower first week activity 00:24:26.509 --> 00:24:29.009 and a higher churn the why is a habit formation 00:24:29.009 --> 00:24:31.670 so let's think about this question right so let's 00:24:31.670 --> 00:24:35.430 just say you joined this streaming uh company 00:24:35.430 --> 00:24:41.630 does the sign up discount that you got lead to 00:24:41.630 --> 00:24:44.150 a lower first week activity and higher churn 00:24:44.150 --> 00:24:46.329 so if they received a discount while joining 00:24:46.329 --> 00:24:53.539 did they get um did they churn more So if they 00:24:53.539 --> 00:24:56.460 got 40 % discount versus maybe 0 % discount, 00:24:56.519 --> 00:24:59.599 did the 40 % discount person churn more than 00:24:59.599 --> 00:25:02.900 the 0 % discount? Probably because the person 00:25:02.900 --> 00:25:05.900 who's paid more wants to stay and try to make 00:25:05.900 --> 00:25:09.059 his money's worth, right? Or yeah, and obviously 00:25:09.059 --> 00:25:11.559 they probably engage with the app as well less. 00:25:12.079 --> 00:25:14.359 I think usually that kind of behavior I engage 00:25:14.359 --> 00:25:17.619 with, right? If I pay more, I'm inclined to stay 00:25:17.619 --> 00:25:19.660 more because, well, I want to get my money's 00:25:19.660 --> 00:25:23.480 worth. um so that is more of a habit formation 00:25:23.480 --> 00:25:26.660 thing that's what the ai comes up with third 00:25:26.660 --> 00:25:31.059 question do week two escalations predict churn 00:25:31.059 --> 00:25:35.099 so this customer has been there on in week two 00:25:35.099 --> 00:25:38.099 they try to do a lot of support ticket escalations 00:25:38.099 --> 00:25:42.259 does that help predict churn so they create a 00:25:42.259 --> 00:25:44.619 lot of support tickets and if those tickets have 00:25:44.619 --> 00:25:47.000 been escalated does that mean that they are going 00:25:47.000 --> 00:25:50.549 to churn Why? Because the early customers, they 00:25:50.549 --> 00:25:55.150 are early experience pain. So that they want 00:25:55.150 --> 00:25:58.130 to try to avoid. So that was the third question. 00:25:58.710 --> 00:26:02.549 Now the fourth question is any region and device 00:26:02.549 --> 00:26:06.049 combos with short tenure? So you want to see 00:26:06.049 --> 00:26:11.950 if maybe people from Northeast United States, 00:26:11.970 --> 00:26:20.029 for example, are they using mobile more but they 00:26:20.029 --> 00:26:24.529 also stay on the platform less they use the streaming 00:26:24.529 --> 00:26:27.589 device very less right i mean the streaming subscription 00:26:27.589 --> 00:26:29.789 very less so that's something you'd want to see 00:26:29.789 --> 00:26:36.309 so why the why behind this is it is a user interface 00:26:36.309 --> 00:26:40.609 parity or localization right is it getting skewed 00:26:40.609 --> 00:26:44.130 in one area or not is is uh the geographic area 00:26:44.130 --> 00:26:46.809 playing a part that's what you want to see here 00:26:46.809 --> 00:26:50.049 and uh so another question it came up with was 00:26:50.049 --> 00:26:54.250 is the churn non -linear versus the first week 00:26:54.250 --> 00:26:57.430 activity so let's just say you know you're trying 00:26:57.430 --> 00:27:00.450 to churn right is it you know at all proportional 00:27:00.450 --> 00:27:03.650 with how much you're interacting in the first 00:27:03.650 --> 00:27:05.730 week so is the first week activity depending 00:27:05.730 --> 00:27:08.549 on i mean is the first week activity affecting 00:27:08.549 --> 00:27:12.609 churn like the more you interact meaning the 00:27:12.609 --> 00:27:14.779 less you churn or something like that right or 00:27:14.779 --> 00:27:19.319 is it not related as much not related alone with 00:27:19.319 --> 00:27:21.880 the first week activity so maybe it's related 00:27:21.880 --> 00:27:25.680 but like not in a linear way you know this will 00:27:25.680 --> 00:27:30.059 help you uh the y here is like an activation 00:27:30.059 --> 00:27:34.339 threshold so you can obviously simplify this 00:27:34.339 --> 00:27:38.279 language and everything but that's how uh the 00:27:38.279 --> 00:27:40.619 kind of questions the tool will help you generate 00:27:40.619 --> 00:27:43.240 and it'll give you some data health notes also 00:27:44.039 --> 00:27:47.680 give you negative 10 -year days, inconsistent 00:27:47.680 --> 00:27:51.440 region labels, and 12 % missing data, missing 00:27:51.440 --> 00:27:54.039 discount applied. So this is more of a data health 00:27:54.039 --> 00:27:58.059 check for, maybe you can think of it like data 00:27:58.059 --> 00:28:01.819 consistency or, you know, you want to have some 00:28:01.819 --> 00:28:04.279 consistent data. So if you see some 10 -year 00:28:04.279 --> 00:28:06.539 days with a negative, it means something is up, 00:28:06.700 --> 00:28:09.200 right? You wouldn't expect these to be negative. 00:28:09.920 --> 00:28:14.819 Or the region labels. Where the person is from, 00:28:15.000 --> 00:28:18.599 you would not want to see different spellings, 00:28:18.599 --> 00:28:22.279 for example, different labels. So that is affecting 00:28:22.279 --> 00:28:25.420 something. And the other thing is discount applied. 00:28:26.599 --> 00:28:29.299 If there's any missing data or something, that 00:28:29.299 --> 00:28:31.500 is what it is telling you. And then finally, 00:28:31.539 --> 00:28:34.140 an export of one page brief into paste into Slack 00:28:34.140 --> 00:28:37.960 and align the team. So these are the three steps 00:28:37.960 --> 00:28:41.809 for that export. So here are some choices. the 00:28:41.809 --> 00:28:44.750 product or engineering can make put the objective 00:28:44.750 --> 00:28:48.009 first focus second and then the questions come 00:28:48.009 --> 00:28:51.009 in so business objective goes first what do you 00:28:51.009 --> 00:28:54.049 want to focus on columns wise and stuff and then 00:28:54.049 --> 00:28:58.049 what questions you want to ask then pre -compute 00:28:58.049 --> 00:29:01.430 small statistics so that the model reasons more 00:29:01.430 --> 00:29:06.269 and you know rummage is less what does rummage 00:29:06.269 --> 00:29:10.400 mean really it means that it's not it's not going 00:29:10.400 --> 00:29:14.680 systematically, right? So you want the model 00:29:14.680 --> 00:29:18.440 to go systematically. So you want it to reason 00:29:18.440 --> 00:29:22.319 more. So small statistic needs to be computed, 00:29:22.460 --> 00:29:25.799 but also monitor these guardrails. Guardrails 00:29:25.799 --> 00:29:27.680 are like, you know, one thing's going up, you 00:29:27.680 --> 00:29:28.980 don't want other things to go down and stuff. 00:29:29.119 --> 00:29:31.759 So guardrail is one thing's going up, you don't 00:29:31.759 --> 00:29:33.559 want other things going down. So guardrail is 00:29:33.559 --> 00:29:38.359 like, you matrix like churn, bounce rate. So 00:29:38.359 --> 00:29:42.210 any kind of, uh fly any kind of leakage is there 00:29:42.210 --> 00:29:45.950 or uh data leakage or messy labels are there 00:29:45.950 --> 00:29:50.269 those should be you know addressed keep the copy 00:29:50.269 --> 00:29:54.650 simple and reduce any blank box anxiety with 00:29:54.650 --> 00:29:56.930 examples so what does this mean people freeze 00:29:56.930 --> 00:29:59.490 basically when they uh let's look at what does 00:29:59.490 --> 00:30:01.710 keep copy simple mean so when you're writing 00:30:01.710 --> 00:30:05.170 the text the users see that the you know which 00:30:05.170 --> 00:30:07.809 includes labels instructions button or button 00:30:07.809 --> 00:30:11.859 text don't get any don't get clever or verbose 00:30:11.859 --> 00:30:15.400 the instead of articulate your analytical objective 00:30:15.400 --> 00:30:18.859 here just say what's your goal what do you want 00:30:18.859 --> 00:30:22.859 to find out right just simple words less thinking 00:30:22.859 --> 00:30:26.220 and it's a smoother flow so that's simple uh 00:30:26.220 --> 00:30:30.019 copy and blank box anxiety means like people 00:30:30.019 --> 00:30:33.220 freeze when they see a blank input field with 00:30:33.220 --> 00:30:37.240 no guidance so example here would be if the tool 00:30:37.240 --> 00:30:40.259 asks what's your objective and leaves a giant 00:30:40.259 --> 00:30:43.160 empty box a user might not know what to type 00:30:43.160 --> 00:30:47.279 this is called blank box anxiety the stress basically 00:30:47.279 --> 00:30:49.779 you're starting from nothing and with examples 00:30:49.779 --> 00:30:54.140 here means so you remove that anxiety by showing 00:30:54.140 --> 00:30:59.319 an example or two right in the box or next to 00:30:59.319 --> 00:31:03.660 it yeah some example placeholders example it 00:31:03.660 --> 00:31:10.319 would be explain the dip in the week of 7 .06 00:31:10.319 --> 00:31:15.819 to 7 .12 right that was a week or explain like 00:31:15.819 --> 00:31:19.119 a march sales dip or something that or it could 00:31:19.119 --> 00:31:22.160 be fine churn drivers for monthly plan now basically 00:31:22.160 --> 00:31:25.559 the user sees how how to phrase their input and 00:31:25.559 --> 00:31:29.240 they can just copy the style to summarize it 00:31:29.240 --> 00:31:32.359 basically means use short and plain words in 00:31:32.359 --> 00:31:35.940 your ui copy And whenever you ask users to type 00:31:35.940 --> 00:31:39.940 something into a box, show them an example so 00:31:39.940 --> 00:31:43.000 that they don't freeze up. Always prompt them 00:31:43.000 --> 00:31:46.480 so they're not left guessing. And to now go over 00:31:46.480 --> 00:31:49.319 guardrails, what it means is like when you're 00:31:49.319 --> 00:31:53.339 building or using an AI EDA helper, like an exploratory 00:31:53.339 --> 00:31:56.279 data analysis helper for AI, you don't want it 00:31:56.279 --> 00:31:59.059 to silently accept everything. You want it to 00:31:59.059 --> 00:32:01.079 flag potential issues that it sees in a data 00:32:01.079 --> 00:32:03.859 set, but in a friendly guiding way. but not in 00:32:03.859 --> 00:32:07.059 a scary error message. Here are some two common 00:32:07.059 --> 00:32:10.160 issues, right? One is on the leakage side. That 00:32:10.160 --> 00:32:12.920 could mean that the data that cheats by including 00:32:12.920 --> 00:32:17.000 information, you wouldn't realistically know 00:32:17.000 --> 00:32:22.519 at the prediction time. So example here is, let's 00:32:22.519 --> 00:32:24.660 say you're trying to predict churn at a signup, 00:32:24.740 --> 00:32:27.579 right? But your dataset includes a column called 00:32:27.579 --> 00:32:31.079 churn in 30 days. That's leakage. It gives away 00:32:31.079 --> 00:32:34.480 the answer. the guardrail is a to the tool that 00:32:34.480 --> 00:32:39.220 the tool can warn um with the warning heads up 00:32:39.220 --> 00:32:42.599 column churned in 30 days looks like an outcome 00:32:42.599 --> 00:32:46.339 variable using it might cause leakage so that's 00:32:46.339 --> 00:32:53.039 what is a leakage right you don't want um your 00:32:53.039 --> 00:32:55.299 output variable to be essentially included in 00:32:55.299 --> 00:32:58.680 the input which is called leakage which would 00:32:58.680 --> 00:33:01.440 mean your dataset is overtrained or overfitted 00:33:01.440 --> 00:33:05.420 in machine learning terms. Now, messy labels 00:33:05.420 --> 00:33:08.380 is the other thing which happens. Definition 00:33:08.380 --> 00:33:13.000 here would be, let's just say your categorical 00:33:13.000 --> 00:33:17.500 values that are inconsistent, redundant, or even 00:33:17.500 --> 00:33:22.660 unclear. So example, let's take a region column 00:33:22.660 --> 00:33:28.660 with values of US, US. let's say you have a region 00:33:28.660 --> 00:33:32.339 column with values us and then another one with 00:33:32.339 --> 00:33:37.000 u .s. and another one with usa and another one 00:33:37.000 --> 00:33:39.720 with north america and all are mixed together 00:33:39.720 --> 00:33:43.759 right the tool can say as a guardrail notice 00:33:43.759 --> 00:33:49.980 region has overlapping categories us u .s. and 00:33:49.980 --> 00:33:53.240 usa you may want to clean this and it will give 00:33:53.240 --> 00:33:55.180 you like friendly tips right like that so if 00:33:55.180 --> 00:33:59.359 the tool says Error invalidator, that's a very 00:33:59.359 --> 00:34:01.460 vague message and users can freeze about it. 00:34:01.819 --> 00:34:04.059 They'll be like, what does that invalidator mean, 00:34:04.180 --> 00:34:06.859 right? Users can just give helpful tips, be helpful, 00:34:07.099 --> 00:34:10.679 right? Or if it says quick tip, you might want 00:34:10.679 --> 00:34:14.300 to check X. Users might feel supported instead 00:34:14.300 --> 00:34:17.599 of judged. You might want to check column X or 00:34:17.599 --> 00:34:20.940 value X in the column or whatever. Quick tip. 00:34:21.800 --> 00:34:25.880 That's a very nice way to say it. basically the 00:34:25.880 --> 00:34:29.219 tone matters small emojis or short wording let 00:34:29.219 --> 00:34:31.619 your ai be your friend so it should be no jargon 00:34:31.619 --> 00:34:34.519 really so what this essentially means is like 00:34:34.519 --> 00:34:36.480 you know you're building soft card rails into 00:34:36.480 --> 00:34:39.579 the tool that gently warns users about common 00:34:39.579 --> 00:34:42.039 pitfalls like leakage or messy category labels 00:34:42.039 --> 00:34:45.000 without breaking their flow now i just wanted 00:34:45.000 --> 00:34:47.179 to again reiterate about this other point about 00:34:47.179 --> 00:34:50.260 pre -computing small stats so the model reasons 00:34:50.260 --> 00:34:53.489 more and rummages less. So this one's about helping 00:34:53.489 --> 00:34:56.289 the AI spend its brain power wisely. So let's 00:34:56.289 --> 00:34:59.210 break it down. So the problem is when you don't 00:34:59.210 --> 00:35:02.230 pre -compute, if you just throw the data set, 00:35:02.329 --> 00:35:06.309 raw data set into a model and say, ask 10 smart 00:35:06.309 --> 00:35:10.530 EDA questions, the model has to pass every column 00:35:10.530 --> 00:35:13.510 from scratch. It may miss obvious things like 00:35:13.510 --> 00:35:15.610 columns which are numeric versus categorical 00:35:15.610 --> 00:35:19.869 and it wastes tokens and time. ai tokens and 00:35:19.869 --> 00:35:22.710 obviously time to rummage the data so to speak 00:35:22.710 --> 00:35:25.969 through the raw csv it sometimes gives generic 00:35:25.969 --> 00:35:28.230 or irrelevant questions and rummage just means 00:35:28.230 --> 00:35:31.150 like it goes about in a messy way very generic 00:35:31.150 --> 00:35:32.789 or relevant questions so you don't want that 00:35:32.789 --> 00:35:35.230 to happen it's like asking a human analyst to 00:35:35.230 --> 00:35:38.530 read a 500 page log file line by line instead 00:35:38.530 --> 00:35:41.769 of giving them a one page summary first so give 00:35:41.769 --> 00:35:44.389 some context context is very important essentially 00:35:44.389 --> 00:35:47.250 right one page summary is essentially a context 00:35:47.840 --> 00:35:50.300 And I'll talk about context engineering in another 00:35:50.300 --> 00:35:53.920 episode, but brief thing is just like to give 00:35:53.920 --> 00:35:57.059 context, right? If you want somebody to do the 00:35:57.059 --> 00:36:00.380 job correctly, help them understand what they're 00:36:00.380 --> 00:36:04.480 trying to do. Here's a simple fix to pre -compute 00:36:04.480 --> 00:36:08.099 small stats. Before asking your model to generate 00:36:08.099 --> 00:36:11.139 questions, you or your app should run a quick 00:36:11.139 --> 00:36:14.440 Pandas summary. For each column, compute things 00:36:14.440 --> 00:36:20.840 like type numeric. categorical text date, count 00:36:20.840 --> 00:36:24.559 of unique values, percent missing values. So 00:36:24.559 --> 00:36:27.159 get a count of unique values, see how much percent 00:36:27.159 --> 00:36:28.840 of missing values are there, what are the data 00:36:28.840 --> 00:36:31.800 types are there, what is the mean, median, and 00:36:31.800 --> 00:36:35.619 standard deviation for the numeric columns. You 00:36:35.619 --> 00:36:37.199 can use a lot of Pandas functions available. 00:36:37.539 --> 00:36:40.159 So I think describe or something does it. And 00:36:40.159 --> 00:36:44.460 info will give you the type. Or a Pandas profiling 00:36:44.460 --> 00:36:47.070 report also. is another library which will give 00:36:47.070 --> 00:36:50.150 you all these answers and in this and and there's 00:36:50.150 --> 00:36:52.409 a couple of other ones like top three most common 00:36:52.409 --> 00:36:55.610 values for categoricals or a min max date for 00:36:55.610 --> 00:36:58.789 time columns so instead of dumping the full csv 00:36:58.789 --> 00:37:02.070 you give the model this compact cheat sheet and 00:37:02.070 --> 00:37:04.190 i'll link uh and i'll give you that in the show 00:37:04.190 --> 00:37:06.469 notes as well why it helps is because basically 00:37:06.469 --> 00:37:09.969 the model can reason about the data and plan 00:37:09.969 --> 00:37:14.239 type it would be like hmm plant type only has 00:37:14.239 --> 00:37:16.519 three categories maybe i can compare churn across 00:37:16.519 --> 00:37:22.340 them so this kind of context will enable the 00:37:22.340 --> 00:37:24.300 model to think smarter right it'll say like oh 00:37:24.300 --> 00:37:26.239 i only have three categories in the plant type 00:37:26.239 --> 00:37:31.099 maybe i can compare churn so it doesn't really 00:37:31.099 --> 00:37:34.599 waste tokens or hallucinate right ai models are 00:37:34.599 --> 00:37:36.599 known to hallucinate a lot So you don't want 00:37:36.599 --> 00:37:38.940 to waste your AI tokens there. Because AI tokens 00:37:38.940 --> 00:37:41.340 is obviously very expensive. So just use it wisely. 00:37:42.840 --> 00:37:46.039 Let it come up with smarter questions. Let it 00:37:46.039 --> 00:37:48.960 say, maybe this is a text column. Basically what 00:37:48.960 --> 00:37:51.159 happens in this case is output gets sharper, 00:37:51.380 --> 00:37:55.679 faster and cheaper. So essentially this whole 00:37:55.679 --> 00:37:57.400 thing means like you'd want to do a lightweight 00:37:57.400 --> 00:38:00.260 summary of the data set first. Then pass that 00:38:00.260 --> 00:38:03.500 summary to AI. That way the AI spends more energy 00:38:03.500 --> 00:38:06.630 on reasoning. and less on digging through ROROs. 00:38:06.769 --> 00:38:09.929 So this is essentially like a message for product 00:38:09.929 --> 00:38:12.469 and engineering. And finally, one last message 00:38:12.469 --> 00:38:15.769 for product and engineering is nothing leaves 00:38:15.769 --> 00:38:18.469 your session unless you opt in. So what does 00:38:18.469 --> 00:38:21.690 that mean? Here's the concern, right? It's a 00:38:21.690 --> 00:38:24.289 trust signal line. Here's what it means in plain 00:38:24.289 --> 00:38:27.949 words. The concern is basically when people upload 00:38:27.949 --> 00:38:32.130 a dataset into your AI and EDA tool, they immediately 00:38:32.130 --> 00:38:35.960 wonder, Is this being stored somewhere? Is my 00:38:35.960 --> 00:38:39.739 client's data safe? Will this end up draining 00:38:39.739 --> 00:38:42.300 someone's model? That's a very valid concern. 00:38:42.480 --> 00:38:45.420 So you need to make sure that you are not opted 00:38:45.420 --> 00:38:50.099 in if you don't want to share your data. So there's 00:38:50.099 --> 00:38:55.019 a way to check that in different tools. In chat, 00:38:55.059 --> 00:39:00.989 GPT is pretty easy to do that. People have very 00:39:00.989 --> 00:39:03.610 valid reasons to have this kind of privacy concerns, 00:39:03.849 --> 00:39:07.170 what we call as a privacy anxiety. You don't 00:39:07.170 --> 00:39:11.150 want the data to be used by other people because 00:39:11.150 --> 00:39:13.510 maybe it's confidential. You don't want to end 00:39:13.510 --> 00:39:18.570 up trading someone's model, obviously. Is this 00:39:18.570 --> 00:39:21.170 being stored somewhere? So that are very valid 00:39:21.170 --> 00:39:23.329 concerns, right? So just make sure you're not 00:39:23.329 --> 00:39:27.179 opted in. Here's something you can watch out 00:39:27.179 --> 00:39:29.639 for is you can calm that by setting like a clear 00:39:29.639 --> 00:39:32.920 privacy boundary. By default, all your processing 00:39:32.920 --> 00:39:34.760 happens locally or within the active session. 00:39:35.099 --> 00:39:37.559 But when the user closes the tab, the data is 00:39:37.559 --> 00:39:41.840 gone. Nothing is logged, saved or even reused 00:39:41.840 --> 00:39:45.800 unless the user chooses to export, save or share 00:39:45.800 --> 00:39:50.480 that. Here's how I would frame the copy. So in 00:39:50.480 --> 00:39:53.320 app or in notes, this is how it would go. So 00:39:53.320 --> 00:39:55.780 your file is processed only inside your session. 00:39:56.199 --> 00:39:58.480 Nothing leaves or nothing is stored. This is 00:39:58.480 --> 00:40:00.400 a more privacy -first approach, where if you 00:40:00.400 --> 00:40:03.119 want to save or share your results, that's only 00:40:03.119 --> 00:40:06.380 opting in. So to summarize, here's what my message 00:40:06.380 --> 00:40:08.619 to product and engineering is. I'll start with 00:40:08.619 --> 00:40:11.320 the last one first. Your tool won't quietly send 00:40:11.320 --> 00:40:14.980 data to your servers, store logs or retrain models 00:40:14.980 --> 00:40:17.780 with their data. Users will stay in control. 00:40:18.269 --> 00:40:20.170 Because sharing is a choice and it's not the 00:40:20.170 --> 00:40:22.869 default. Building soft guardrails into the tool 00:40:22.869 --> 00:40:25.590 that gently warn users about common pitfalls 00:40:25.590 --> 00:40:29.949 like leakage or messy category labels without 00:40:29.949 --> 00:40:33.349 breaking their flow. So build those guardrails 00:40:33.349 --> 00:40:35.989 into the tool that warns users about pitfalls 00:40:35.989 --> 00:40:40.889 like leakage or messy category labels. Be supportive 00:40:40.889 --> 00:40:47.400 about it as well. Next point. Use short and plain 00:40:47.400 --> 00:40:50.559 words in your user interface copy. And whenever 00:40:50.559 --> 00:40:53.380 you ask users to type something into a box, show 00:40:53.380 --> 00:40:56.539 them an example of how to do it so that they 00:40:56.539 --> 00:40:59.539 don't freeze up. We want to make everything easy. 00:40:59.739 --> 00:41:04.000 And finally, put your objective first, focus 00:41:04.000 --> 00:41:05.860 second, and then question. That's really the 00:41:05.860 --> 00:41:08.280 first point here. Objective first, focus second, 00:41:08.340 --> 00:41:11.119 and questions last. Here's the numbers I track 00:41:11.119 --> 00:41:16.349 which are, you know, more useful than cool. percent 00:41:16.349 --> 00:41:20.250 of sessions with an objective like 80 percent 00:41:20.250 --> 00:41:23.170 greater than 80 percent of the target variable 00:41:23.170 --> 00:41:26.110 right percent of useful questions time to brief 00:41:26.110 --> 00:41:31.289 like ideally under three minutes seven day return 00:41:31.289 --> 00:41:33.570 so more on the habit side like the user activity 00:41:33.570 --> 00:41:36.289 side and the real world use is this is how i 00:41:36.289 --> 00:41:40.590 usually try to work with data set i prefer using 00:41:40.590 --> 00:41:44.130 churn also as like a Churn rate over a period 00:41:44.130 --> 00:41:46.329 of 30 days. That's my useful metric. That's why 00:41:46.329 --> 00:41:52.889 I kept talking about it. Retention over a three 00:41:52.889 --> 00:41:54.969 -month window, a six -month window, a yearly 00:41:54.969 --> 00:41:56.929 window. They're very common metrics to look at, 00:41:57.010 --> 00:42:00.170 but these are very useful also. Better than your 00:42:00.170 --> 00:42:04.309 vanity metrics like followers or subscribers. 00:42:04.949 --> 00:42:08.449 That's very vanity metrics. Number of users, 00:42:08.610 --> 00:42:11.659 for example. Those are not going to help. user 00:42:11.659 --> 00:42:14.579 growth is what really helps. And finally, let's 00:42:14.579 --> 00:42:16.719 come to the listener checklist. I'll be also 00:42:16.719 --> 00:42:18.760 adding this in the show notes, but here's what 00:42:18.760 --> 00:42:22.920 I just want to read out first. This is a follow 00:42:22.920 --> 00:42:25.340 -up to the quiz you listened to earlier. So let's 00:42:25.340 --> 00:42:27.579 go with this checklist here. This is the third 00:42:27.579 --> 00:42:30.780 key item I wanted to discuss. First, you write 00:42:30.780 --> 00:42:34.920 the objective in one line. Like example would 00:42:34.920 --> 00:42:39.489 be reduce x by y in z time frame for whatever 00:42:39.489 --> 00:42:42.710 the segment is um pick five to eight focus columns 00:42:42.710 --> 00:42:46.849 or accept the tools hints generate 10 questions 00:42:46.849 --> 00:42:50.409 plus one line of why it matters so this is like 00:42:50.409 --> 00:42:53.250 a prompt you are using essentially four run quick 00:42:53.250 --> 00:42:58.250 data health checks nulls odd values naming export 00:42:58.250 --> 00:43:02.889 a one page markdown brief number five number 00:43:02.889 --> 00:43:06.309 six share it in slack notion jira and tag owners 00:43:06.309 --> 00:43:10.369 number seven Run two to three quick analysis 00:43:10.369 --> 00:43:14.590 today, not 10. Log what you learned, plus the 00:43:14.590 --> 00:43:17.989 next decision. Nine, repeat tomorrow with tighter 00:43:17.989 --> 00:43:22.710 focus. And tell me which step saved you the most 00:43:22.710 --> 00:43:26.070 time in the sub stack comments. So I'll list 00:43:26.070 --> 00:43:28.889 this checklist in the show notes so you can have 00:43:28.889 --> 00:43:30.670 it. All you need to do is just don't overthink 00:43:30.670 --> 00:43:32.550 it. Just run through these nine steps and you'll 00:43:32.550 --> 00:43:35.570 move faster than 90 % of people who dropped into 00:43:35.570 --> 00:43:38.400 a data set. right i'd love to hear from you which 00:43:38.400 --> 00:43:41.860 step actually saved you the most time so again 00:43:41.860 --> 00:43:43.820 just drop your answer in the comment section 00:43:43.820 --> 00:43:47.559 i'll link the thread for you in substack and 00:43:47.559 --> 00:43:50.539 i read them all like i said i'll share a few 00:43:50.539 --> 00:43:53.239 next week if you'd want so we can all learn from 00:43:53.239 --> 00:43:56.119 each other so what to steal from this show and 00:43:56.119 --> 00:43:58.599 maybe think of it like an action plan for you 00:43:58.599 --> 00:44:03.000 points start every analysis with the decision 00:44:03.000 --> 00:44:06.400 limiting yourself to five to eight columns Always 00:44:06.400 --> 00:44:09.820 ship a one -page brief. Track usefulness and 00:44:09.820 --> 00:44:13.820 not vanity metrics. And finally, make the first 00:44:13.820 --> 00:44:16.960 90 seconds feel like progress. All right? Before 00:44:16.960 --> 00:44:20.199 you go, I just wanted to let you know that I 00:44:20.199 --> 00:44:23.860 record with Riverside. I'm an affiliate partner 00:44:23.860 --> 00:44:27.139 of Riverside. They helped me record this episode. 00:44:27.739 --> 00:44:31.900 And they have amazing audio quality. So that's 00:44:31.900 --> 00:44:33.699 something I just wanted to recommend if you're 00:44:33.699 --> 00:44:36.380 ever, you know, deciding to do your own podcast. 00:44:37.219 --> 00:44:40.760 This is a great platform to record because their 00:44:40.760 --> 00:44:43.780 editing features are amazing. Their AI editing 00:44:43.780 --> 00:44:46.699 is, I think, next level. Helps you remove all 00:44:46.699 --> 00:44:50.539 awkward pauses and gives your sound a strong 00:44:50.539 --> 00:44:55.119 voice. It gives you a strong voice. So I would 00:44:55.119 --> 00:44:58.119 highly recommend Riverside as the reason why 00:44:58.119 --> 00:45:00.480 I chose to... be an affiliate partner because 00:45:00.480 --> 00:45:03.940 I actually use the product and I love it. So 00:45:03.940 --> 00:45:08.460 link to join their platform is in the show notes. 00:45:09.119 --> 00:45:12.139 And again, I'm an affiliate partner. So if you 00:45:12.139 --> 00:45:15.739 make any purchase, I may make a small commission 00:45:15.739 --> 00:45:20.760 from it. I also host an rss .com. Also, I'm an 00:45:20.760 --> 00:45:25.239 affiliate partner of RSS because they have an 00:45:25.239 --> 00:45:29.119 amazing distribution channel. They are able to 00:45:29.119 --> 00:45:33.840 publish your episodes from RSS into Spotify, 00:45:34.159 --> 00:45:38.840 Apple Podcasts, Amazon Music, Audible, Deezer, 00:45:39.139 --> 00:45:44.579 Pandora, iHeartRadio, and even lets you publish 00:45:44.579 --> 00:45:47.260 your episode with one button into YouTube as 00:45:47.260 --> 00:45:50.659 well. Oh, did I forget to mention that you can 00:45:50.659 --> 00:45:55.559 start making money by hosting on RSS .com even 00:45:55.559 --> 00:45:58.559 if you have just 10 downloads a month. So that's 00:45:58.559 --> 00:46:00.559 all you need to make money. 10 downloads a month. 00:46:01.900 --> 00:46:04.079 It's a new feature they started and it's amazing. 00:46:04.639 --> 00:46:07.639 So again, link to joinrss .com will be in the 00:46:07.639 --> 00:46:12.300 show notes. I'm also using Cider .ai to help 00:46:12.300 --> 00:46:17.079 with my research. So if you are interested in 00:46:17.079 --> 00:46:21.780 any research for your content or for in general, 00:46:21.920 --> 00:46:24.280 you want to learn about anything, want to read 00:46:24.280 --> 00:46:29.880 research papers or any kind of thing to improve 00:46:29.880 --> 00:46:32.820 your knowledge, cider .ai is your platform to 00:46:32.820 --> 00:46:35.539 use. So again, I'm an affiliate partner and I'd 00:46:35.539 --> 00:46:40.139 love if you're able to support the show by clicking 00:46:40.139 --> 00:46:43.599 on my affiliate link and joining cider .ai. So 00:46:43.599 --> 00:46:47.599 that's all I have, but take the quiz, drop your 00:46:47.599 --> 00:46:49.579 story in the sub stack comments, grab the checklist, 00:46:49.900 --> 00:46:55.989 share this with your friend. With that one friend 00:46:55.989 --> 00:46:59.829 who's drowning in CSVs or Excels. Show notes 00:46:59.829 --> 00:47:02.349 have everything. Substack discussion link is 00:47:02.349 --> 00:47:06.050 also included. And a thread is also there for 00:47:06.050 --> 00:47:09.130 you to join on Substack. So see you next week. 00:47:10.070 --> 00:47:13.949 Thanks again. And this is Mukundan from the Data 00:47:13.949 --> 00:47:15.869 and AI with Mukundan show. Hey, it's Mukundan. 00:47:16.010 --> 00:47:18.570 If this episode helped you, two tiny favors that 00:47:18.570 --> 00:47:21.250 make a huge difference. Rate the show five stars. 00:47:21.900 --> 00:47:24.659 On Spotify, you can just open the show page and 00:47:24.659 --> 00:47:27.960 tap the star button. On Apple Podcasts, scroll 00:47:27.960 --> 00:47:30.340 to the bottom of the show page and tap five stars. 00:47:31.800 --> 00:47:34.820 Also, leave a one -to -line review on Apple Podcasts. 00:47:35.719 --> 00:47:38.280 Tell me one takeaway or which checklist step 00:47:38.280 --> 00:47:41.079 saved you time. I read every single one and it 00:47:41.079 --> 00:47:43.880 helps more people find the show. Thank you. Your 00:47:43.880 --> 00:47:46.599 rating and review genuinely help this reach the 00:47:46.599 --> 00:47:49.440 next analyst who's staring at a messy CSV. See 00:47:49.440 --> 00:47:49.900 you next week.