AWS Builder Center
The 72-Hour Clock: Building an AI Agent to Save Farmers' Insurance Claims on AWS

The 72-Hour Clock: Building an AI Agent to Save Farmers' Insurance Claims on AWS

Farmers have just 72 hours to report crop damage under PMFBY, and many valid claims are rejected on technicalities like name mismatches. Here's how I built Fasal Rakshak, a voice-first AI agent, using AWS Amplify, S3, and DynamoDB during the AWS India Builder Tour Hackathon.

I am developer who love building and shipping

The 72-Hour Clock: I Built an AI Agent on AWS to Save Farmers' Insurance Claims

Picture a farmer in Madhya Pradesh. He has spent four months on his wheat, and one night a hailstorm flattens it. By morning the crop is gone, and a clock has started.
Under India's crop insurance scheme, PMFBY, he has 72 hours to report the loss. Miss that window and the claim can be rejected before a single person looks at the damage. Many farmers never learn the window exists. The ones who do still have to deal with English forms, document requirements, and a process that wasn't designed around their day.
During the AWS India Builder Tour Hackathon, I decided to build something for that clock. I called it Fasal Rakshak, which means "crop protector."


The problem nobody talks about

When I started reading about PMFBY claims, I expected the hard part to be proving the damage. It usually isn't.
A lot of claims fail on technicalities:
  • The deadline. The 72-hour reporting window is strict. Farmers who don't know about it, or can't report in time, get auto-rejected before assessment even begins.
  • The name mismatch. Aadhaar says "Ram Kumar" and the bank account says "Ram Kumar Sharma." It's the same man with the same field and the same loss, but the records don't match and the claim stalls.
  • The language barrier. The forms assume you're comfortable with formal English, and plenty of farmers are not.
None of these say the loss wasn't real. They say the paperwork wasn't perfect. That's the part I wanted to fix, because software is good at exactly this kind of thing.

The idea: don't build a form, build an agent

The easy version of this project is a digital claim form. I didn't want that, because a form still expects the farmer to do all the work.
I wanted an agent that works around the form, so the farmer only has to say what happened. Fasal Rakshak does four things.

1. It listens first

The farmer speaks in their own language. Something like "meri fasal ole se kharab ho gayi" ("my crop was ruined by hail") is enough. The AI turns that into a structured claim and maps it to PMFBY standards, so nobody has to wrestle with English forms.


2. It fixes errors before they become rejections

This is my favorite part. When the agent notices Aadhaar and bank details don't match, it doesn't just flag an error. It proposes a specific fix, and the farmer taps accept. The claim updates in place. The mismatch never reaches the reviewer.

3. It protects the deadline

Every report gets a geotag and an immutable timestamp. If anyone ever disputes when the farmer reported, there's a clear record. The app also shows a live countdown, so the 72 hours stop being an invisible threat.

4. It fights back after a rejection

If a claim does get rejected, the agent drafts a grievance appeal that cites the relevant clause. Most people can't write that kind of appeal without help, and it's the step where many farmers simply give up.


How I used AWS

I had four days and no interest in babysitting servers, so I picked services that stay out of the way.
AWS Amplify hosts the app. I pushed to the branch and got a deployment on a real public URL. For a hackathon that mattered more than I expected, because I could test the app the way a user would while I was still building it, instead of trusting  localhost .
Amazon S3 stores the photo evidence. Photos of damaged crops are the core of a claim, so I wanted them stored somewhere durable and treated as a permanent record instead of something that could quietly change later.
Amazon DynamoDB stores the claims. Each claim is a self-contained record with a status, a timestamp, and a trail of what the agent did. That fits a NoSQL model naturally, and I didn't have to plan for scale up front.
For the AI layer, I used Groq for fast inference. Someone reporting a disaster from a phone shouldn't be staring at a spinner, so speed mattered a lot.

What I learned

This was my first time [ADD YOUR TRUTH HERE: first time deploying on Amplify? first time using DynamoDB or S3 in a real project? first time building a voice-based product?], and it changed how I think about building. [ADD ONE REAL MOMENT: something that broke at 2 a.m., something that surprised you, something you'd do differently.]
A few lessons I'll carry forward:
Start with the smallest painful thing. "Help farmers" is a mission statement. "Don't let a spelling difference cost a farmer his claim" is a product. The second one told me exactly what to build and what to skip.
One working feature beats five half-working ones. Every hour I spent making the mismatch-fix flow work end to end paid off more than an hour spent adding something new.
Managed services buy you time. With Amplify, S3 and DynamoDB handling hosting, storage and data, I spent my four days on the actual problem instead of on infrastructure.

What's still missing

Fasal Rakshak is a hackathon build, so I want to be honest about that.
It needs testing with real farmers and real PMFBY workflows. It needs more languages and dialects. And I suspect the first farmer who tries it will break half my assumptions in five minutes. That's exactly why I want to sit with a few of them and watch.
It's also an assistive tool. It isn't affiliated with the Government of India, and it doesn't replace the official process. It helps people get through that process without losing a valid claim to a technicality.

The bigger picture

The system fails people in small, boring ways: a missed deadline, a spelling variation, a form in the wrong language. Each one looks minor, but together they decide whether a family gets paid after a bad season.
I don't think AI needs to be flashy to matter. Sometimes it just needs to make sure a farmer's voice gets heard before the clock runs out.
Built for the AWS India Builder Tour Hackathon 2026.


Before you post:
  • Fill the two bracketed spots in "What I learned." They matter most for the human feel, because judges can tell a real story from a generic one.
  • Image 2 has a visible glitch. The agent's reply starts with "'ve recorded" (the "I" is missing) and cuts off at "72 hours to". If you can, fix that text or take a fresh screenshot, because a broken sentence in a screenshot can hurt you.
  • Suggested captions for the images: (1) "Fasal Rakshak: an AI agent for PMFBY claims", (2) "A farmer reports hail damage by voice in Hindi", (3) "Voice intake, Aadhaar-bank checks, damage assessment, appeals, and a 72-hour safeguard in one place".
  • Statistics. I left out the 40% and 25% figures from your site, since I can't verify them. Add them only with a source.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article