Saarthi

An AI agent that takes farmers from “What’s my crop worth?” to “Where should I sell?”

Comment

Inspiration

Small and marginal farmers in India routinely lose ₹15,000–30,000 per harvest cycle to opaque mandi pricing, unreliable transport, and zero visibility into who's actually buying nearby. In our research (60+ farmer conversations), the same three complaints kept surfacing: "I don't know the real price," "I don't know who will pay more," and "I can't get a truck on time." Saarthi (सारथी — "charioteer/guide") exists to fix that: a farmer speaks a single sentence about their produce and gets an instant, ranked list of nearby buyers, transporters, and cold storage — with a price comparison and one-tap Call/Navigate actions.

What it does

Saarthi is a multilingual (Hindi/English) agri-logistics platform with four connected roles:

  • Farmers list produce, get AI-ranked buyer/transporter recommendations, request pickups, and track orders.
  • Buyers/Traders discover fresh produce nearby, place orders, and track fulfillment.
  • Transporters see nearby logistics jobs, accept bookings, and get turn-by-turn navigation.
  • Cold storage operators manage capacity and connect it into the discovery flow.

A conversational assistant ("Sarthi Didi"), powered by Gemini, extracts intent from natural-language input (e.g. "I have 80kg tomatoes ready tomorrow") and returns a ranked comparison: best immediate sale vs. best price if you can wait, distance-weighted, with a plain-language recommendation in the farmer's own language. Booking flows use live Google Maps routing (with an OpenStreetMap/OSRM fallback so the whole app still works with zero paid API keys), and a mock payment/split-ledger flow demonstrates the full order-to-payout lifecycle.

How we built it

  • Frontend: React 19 + TypeScript on Vite, React Router for the multi-role app shell, Framer Motion/GSAP/Three.js for the landing experience and onboarding.
  • AI: Google's Gemini API (@google/genai) for intent extraction, transport-option generation, and the bilingual support assistant — with graceful mock fallbacks whenever no API key is configured, so the demo never breaks.
  • Maps/logistics: @vis.gl/react-google-maps + Directions API when a key is present; Leaflet + the public OSRM demo router otherwise.
  • Data: Supabase (Postgres) for shared farmer/buyer/transporter registries and self-onboarding; data.gov.in for live mandi commodity prices; OpenWeatherMap for weather-aware advisories.
  • Payments: a mocked Razorpay-shaped checkout with a split-ledger view, ready to swap for real Razorpay keys.
  • Deployment: Vite build shipped to Vercel via GitHub Actions, so every push to a branch gets its own preview and main auto-deploys to production.
  • We built a mock-first architecture throughout: every external dependency (AI, maps, payments, OTP) has a working offline/mock mode, so the product is fully demoable without any paid credentials — critical for a hackathon judging environment.

Challenges we ran into

  • Designing for connectivity- and literacy-constrained users meant voice-first, Hindi-first UX had to be a first-class citizen, not a translation layer bolted on afterward.
  • Keeping a full demo working with zero API keys — every AI/maps/payment integration needed a realistic mock path, which took real engineering discipline (schema-validated mock data, not just hardcoded strings).
  • Reconciling a fast-moving MVP scope with a larger platform vision — mid-build we pivoted parts of the roadmap toward satellite-based crop intelligence (Google's AlphaEarth) and had to keep the working logistics core stable while building an AI advisory layer on top.
  • Multi-role state and routing — one shell serving four very different user roles (farmer, buyer, transporter, cold storage) with a shared session/auth layer required careful separation between the v2 redesign and legacy components while both were in flight.

What we learned

That the "AI feature" farmers actually want isn't a chatbot — it's a fast, trustworthy answer to "who should I sell to, right now, for how much." Voice and language access matters more than UI polish. And building resilience-by-default (mock fallbacks everywhere) turned out to be as valuable for judging/demo day as it will be for real-world rollouts in low-connectivity regions.

What's next for Saarthi

We have been selected in sarvam AI startup program, and will be using the tools api given in Integrating satellite-based crop health signals (Google AlphaEarth) and weather-aware advisory (Gemini) to move from "sell better" to "grow better," real Razorpay settlement, and expanding live mandi price coverage and cold-storage partner density beyond our initial pilot region.

Built With

Share this project:

Updates