Introduction

This is my very first talk, so thanks Reggie, Anna, I think, I hope I'm doing, I'm going to be great.

I'll be talking today about Krishi AI. It's an app that I created for Indian farmers, not just Indian farmers, but farmers. It uses LLMs to answer some basic questions that a farmer might have.

So maybe at the end I can give you a small demo or walk you through.

If you are wondering what that is, that is my specs or requirement document that I created in 2015 which I gave it to one of the freelancers to generate an app for me.

Why Krishi AI: The problem worth solving

Before I tell about Krishi AI and what I did with LLMs, I should tell you why I chose this particular problem and why me and no why why I did not say okay I want to develop Facebook AI dot 2 .0 why I chose Krishi AI and most likely why it was me who chose it these are the people who work in

my farm on a daily wages of around three dollars a day and when I as I said when I worked when I developed an app this was the picture in my eyes and I thought I should do something.

The yield gap between India and Europe

Almost for 20 years or more I have been working in two systems one is back there in India the farmers who work almost very hard on a daily basis earning hand -to -mouth on the other hand we have some European people or European farmers who makes a good living out of professions or farming.

The stark

differences I found was, just to give an example, a tomato farmer in Netherlands has almost 25 to 28 times more yield per acre or per hectare as compared to a farmer in India I'm not talking about 25 % it is 25 times and then I was

digging why this particular difference is Indian farmer lazy do they work less do they have less intelligence most likely not I drill down to this rabbit hole and then I found out most there are many more causes but most likely these

Three root causes: knowledge, processes, and pricing

were my top three causes of why there was a difference it's knowledge they were lacking knowledge they most likely did not follow many processes they were not aware of processes and the price they get for their commodities to maybe

A small example: why plowing in the afternoon matters

just give you an example when you plow your field you should do it in the afternoon or when there is sunlight why I did not know until I was 40 years old

I asked my mom did you know she's been doing farming for almost 50 years she said no because I have never seen anybody doing it in afternoon because we normally do it in summer summers are very hot in India so people either do it in the night or early morning or evenings the thing is you do it in the

afternoon because when you flow the field you have insects especially the larvas and the insect which comes up and only in the afternoon the birds will come and eat it so 80 % of your insect problem is already resolved during flowing a small trick would you know save almost one full pesticide cycle but

of course neither it's been documented nor any traditional people knew about it it's only written in English document which Indian farmer never had access to

processes I mean if I talk broadly we should do a germination test we should do a soil test I have never seen even a single farmer doing it in India so lack of processes I in almost all the people that I knew price difference just from

yesterday's newspaper if you see one ton of onion was sold at $13 and the cost of cultivating this one ton of onions was $17 so basically farmer lost four

dollars and three months of his efforts and then I said me being in Switzerland working for a finance industry I know that you know we need to have processes we need to have knowledge repository working in a software can I do something for them and that's exactly where the first thought of Krishi AI was born in

First attempt (2015–16): the original farming app idea

2015 -16 it was not called Krishi AI then it was called as digital Monday I would tell you what Monday is at later point of time but that was my first attempt to to create an app which will help farmers.

A simple use case: choosing where to sell based on prices

The idea was very simple. For example, you are an apple farmer. You live somewhere between Zurich and Winterthur.

And every day morning, you need to take a call. You need to sell, let's say, one ton of your apples. So do you go to Zurich to sell it, or do you go to Winterthur to sell it?

The decision would be very simple if you know what is the price they're going to give you in Zurich. if it is one franc and if it is going to be one and one franc and ten rapans in winter tour it's no -brainer right so you're going to go to winter tour but

somebody has to tell you what are the prices this morning or two hours ago so that you know I can pack in my truck and take it and that's what my app did it would just tell you the prices of your or real -time prices of your commodity

namely soybean or cotton or some of the commodities that happen to be in India and it will just give you prices.

So idea was that even if it is going to give you 2 % more money for your commodity, I have already done my job.

What it took to build v1.0 (freelancers, cycles, and cost)

So for this app, I spent almost 9 months, more than 200 hours of efforts, 6000 francs of my hobby money and of course I had to switch between 3 different freelancers.

the job was very very tedious it was not very simple there was a small process or the process well I had an idea in my mind I would I will talk to my freelancer on Skype then draw a sketch give it to him wait for two weeks he will give me back the code in APK format I would put it on my mobile test it if I

found a problem you go through these two weeks of cycle one more time but yes nine months and I had my version 1 .0 in production almost 45 ,000 downloads it

Why the first project paused

was doing quite great and then I wanted to do 2 .0 but of course so many constraints because I promised myself I'm not going to charge my farmer at any point of time it would be free for farmers.

I did not find any sponsor or I had no other idea there was no kickstart then or somebody could could sponsor it

and then of course in 2016 priorities of my life also changed we got Sun so my hobby money was moved at some some other place and this app died its natural death.

LLMs change everything—and revive the project

2023 -24 I'm not very sure when chat GPT was born and it completely changed the landscape of how people are gonna code in the future.

I was I was witness of that particular event I saw that tweet from Sam Altman and then I tried.

From classic ML experiments to ChatGPT-driven development

I was aware of AI and ML even before then I did some ML classification in 2020 on my own machine to again classify between healthy plant and a disease plant and then I wanted to integrate within my app but of course as

I said chat GPT changed everything I followed the part through chat GPT 4 .0 I think was a promising LLM you know earlier we could just auto complete stuff you know for I equal to and then it will auto complete your for loop but

But I think with 4 .0, you could really prompt and say, OK, write me a function for hello world or write me a function for addition and blah, blah, blah. And that kept me in a loop of learning. That's where I started learning.

But in 2025, I think with GPT -5, things were completely different when they said agentic code would be integrated within your you know ID is and as I said how do I put

everything together a purpose I mean before that I did all the learnings from Coursera you know this Titanic workshops and you name it and then you know I wanted to have some project where it will fulfill my purpose I do a good learning and feel good about myself and that's where everything came together

my died idea of Krishi AI, my learning of LLMs I wanted to put together and that's where I gave my second try.

Rebuilding with modern tools and lots of iterations

The tools that I used was VS Code, GitHub Copilot, by no mean I'm marketing them but these were my tools. I used Gemini, I I used Android Studio and I had hundreds of throwaway versions.

So you would code, you like it, keep it, if you don't like it, throw it away.

From weeks to minutes: a new development loop

What did things change for me?

As I said, my older loop was around two weeks of time or sometimes three weeks of time just to get one delta change. this data change for me was matter of minutes or hours so I would do almost exactly the same thing I would ask a question I will get the code I will

compile I'll run I'll see for errors and then if I find it good if things were Delta ahead I will go forward otherwise I'll just you know go back to the last

saved checkpoint what LLMs did to me was not only about coding but something even

more than that just to give you a background when I checked with my alpha

LLMs beyond coding: product decisions and architecture

testers or small group of people out there people said that we really like your app but we need some kind of a customization so you know I don't want to put the names of two locations every time i want to keep it saved so i thought okay i need

to go personalized i need to have a login created for people you know so that their preferences can be saved and then i obviously had google login created for them and my alpha testers told me nobody has google account they normally use only mobile numbers and i said if i go for mobile

numbers I researched for two weeks no solution and then I said okay why don't I ask LLM about what is the possible solution and we had a brainstorming sessions of around one one and a half hour and then it told me why don't you

use firebase it has almost up to 100 ,000 free SMS's or you know OTP's that will be generated and authentication will be done via mobile numbers it will have a storage repository for you and at some point of time if you want to send some messages within the

app that facility is also there so it not only just suggested me but it changed the whole architecture of my app and then of course today my app allows google login as well as a mobile login

Debugging hard production issues with LLM help

another interesting fact about llms the way it helped me was of course debugging everybody knows it's a nightmare especially when you have an error and especially this error when it is being reproduced in production but cannot be reproduced on

your machine and that's exactly what happened to me the Monday prices were not being updated in my app when I run the function locally it worked perfectly but only when it was run on Google cloud or cloud it failed it did not update the database and then there was an error it said that you need most this IP address

was blocked the ingress was not allowed so I chose another cloud and of course I faced a similar problem and then lastly LLM told me how about you install it locally and turn it on your own machine and that's what is happening it's there on my Mac this particular function is running until I find the next place now

Where Krishi AI is today

to and it will keep running on my laptop so where are we today now so I started with initially just one simple app or functionality within the app but today

almost 32 ,000 lines of code more than 130 commits and more than 16 production releases I have AI chart integrated within my app I have plant diagnosis is that means I'm using multi model behind the screen of course the money

prices which was my one of this very first thought then I have a knowledge repository and of course I'm trying now with some monetization it's not complete yet but hopefully within next few weeks I'll have this particular functionality

Lessons learned building with LLMs

completed what did I learn from with with my journey for last three to six six months with coding with LLM.

The first draft is easy; refinement takes work

The first draft is very easy. The boilerplates, you give a prompt, it gives you a boilerplate, and then there is a lot of aha moments.

It works. It works quite well within the first attempt.

But if you want to take steps forward, it's not a linear path, it's exponential. So you'll have a lot of hurdles, you'll have to have too many throwaway versions, and then

Practical tips: plan mode, structured outputs, and model experimentation

go back to the last save point I would always suggest I mean there are people on X saying that oh I don't use plan mode but as a beginner or maybe if you have started in last three to six months working with LLMs please use plan mode before you you know ask it to code but that will really save lot of rework for for you.

If you have a database behind, please, please use Pydentic or some kind of structured outputs that will come from LLMs because we all know it's probabilistic. Every time it will suggest 11 columns or sometimes it will suggest 10 columns and then you will keep on popping up errors now and then and you will not know why.

there was there is I mean every day you go on any of the social platform there would be a lot of debates happening which is better model opus or codecs or GPT's and the answer is very simple please experiment find out which is the best shooting for you you know again if you are if you have two buckets you know

you you test with one and develop with another you develop you know test with this one and then develop with another one so please use the different models experiment with them and then i'm pretty sure you will find your best suit for yourself another not suggestion but

Don’t skip Git: versioning and branching save you

another finding that i will say is coding is good but please also learn github git branching that will definitely save you a lot of hustles you you will know where to go back where to fall down to

please use labels or whatever versioning numbers you have otherwise again it's very very different it's difficult to manage the amount of because the amount of code that is generated with LLMs is very vast and it's very easy to be lost

The productivity paradox: more time in flow

maybe one more thing I know I'm on time sorry I mean it's very obvious we think that instead of eight hours of work my work is completed within two hours so what do I do for six hours do I drink coffee do I smoke but what I found that

I have been spending more time in front of the screen with LLMs as compared to before because I mean state of flow more often many times as comparative because there are no frictions you know I ask the question I read what is it writing

what is it trying to do if it doesn't go as per my thought I just you know I try to change the way and it's it's more pleasing it gives me more pleasure and

Conclusion: domain knowledge + low-friction coding can create impact

here we are I started 2015 with this screen on the left hand side and today

is the screen on the right hand side maybe one thing if you want to take away

from today's talk or my experience is you know domain people can really make lot of difference you know the cost of coding or the frictions of coding are really minimalized you know that they stand nowhere if you have interest and

if you have your domain knowledge you can you can build magic thank you