ClearSky

Satellite fire counts in Punjab's Sangrur district fell 93% in three years, and the smoke didn't. Farmers burn paddy stubble because the gap between rice harvest and wheat sowing is a few weeks, and a match is the fastest way to clear a field. ClearSky treats that as a logistics problem, not an awareness problem: remove the reason a farmer reaches for the match.

What it is

A farmer sends one WhatsApp voice note or text with their village, acres and harvest date. The backend books the nearest free baler before the sowing deadline and routes the straw to the industrial buyer that pays best after transport. Balers, buyers and district officers get a web dashboard. It was built in four days by Team Atherion; the full walkthrough is in the blog post.

How it works

• Voice in, booking out. Meta's webhook hits API Gateway and a small Lambda that verifies the signature and queues the message. Amazon Transcribe handles Hindi audio, then an agent calls tools to find and book a baler. The LLM is swappable, with a deterministic rules bot in four languages as fallback.
• No double bookings. A baler-day is a shared resource, so a booking is one DynamoDB transaction with four conditional writes: all succeed or none do.
• Never says "confirmed" early. Until a baler actually accepts, the farmer is told it is an offer, and the prompt forbids claiming otherwise.
• A burn-risk radar. An hourly job scores every unbooked field, so officers can chase the fields most likely to burn first.

My part

I worked on the AI model, testing and the production deploy at clearsky.akkki.tech. The WhatsApp agent and AWS backend were Sanskar Joshi's, the role dashboards Kamran Pathan's, and the pollution-impact view Anushka's.