Inspiration
India has 140 million farming households, yet the average farmer still diagnoses crop disease by asking a neighbour or waiting days for an agronomist. And even when farmers do get a diagnosis, they rarely know which government schemes they're entitled to claim for the loss — or whether to sell their harvest today or wait for a better mandi price tomorrow. That triple gap — between knowledge and access, between entitlement and awareness, between market data and the farmer holding it — is what AgriSaathi was built to close.
What it does
AgriSaathi is an AI-powered crop advisory bot for Indian farmers. A farmer snaps a photo of a diseased leaf — AgriSaathi identifies the disease using a fine-tuned ResNet-50 model, generates a treatment advisory grounded in ICAR knowledge, and translates everything into the farmer's language via IndicTrans2. It also gives matching government scheme eligibility (PM Kisan, PMFBY, KCC) and shows today's mandi price alongside a 10-day Prophet forecast — so the farmer knows not just how to save the crop, but when to sell it.
How we built it
We built entirely on the Databricks platform. Raw data — 54,000 PlantVillage leaf images, ICAR advisory PDFs, myScheme eligibility data, APMC mandi prices from data.gov.in via Auto Loader and scheduled ETL jobs.
The disease classifier is a ResNet-50 fine-tuned on 38 disease classes. The mandi price forecaster is a Prophet model trained on 60 days of APMC data per crop-mandi pair. The advisory layer GENIE Chatbot: Search, retrieved at query time for response generation.
The frontend is a Streamlit app deployed on Databricks Apps, with an AI/BI Lakeview dashboard for extension officers covering disease heatmaps, price trends, and outbreak signals, and a Genie Space for natural language analytics queries over the Delta tables.
Challenges we ran into
Getting IndicTrans2 to serve reliably within Databricks Model Serving latency constraints took significant tuning — the model is large and translation adds a perceptible lag in the end-to-end flow. APMC CSV data from data.gov.in was far noisier than anticipated — deduplication, schema evolution, and cross-mandi normalisation in the Bronze-to-Silver pipeline consumed more time than any other single component. Prophet forecasting worked well on crops with dense price history but degraded on sparse mandi-crop combinations, which required fallback logic. Finally, designing a RAG prompt that produces concise, actionable advice — not verbose LLM output — for a farmer audience required several iterations of system prompt engineering.
Accomplishments that we're proud of
We're proud that the core farmer flow — photo in, disease diagnosed, treatment advised, schemes matched, price forecast shown — works end-to-end in under 10 seconds on a live Databricks Model Serving endpoint. The Prophet-based mandi price forecaster giving a 10-day outlook per crop-mandi pair felt like a genuinely useful signal, not just a demo feature. The outbreak detection logic running on geo-tagged query logs is the kind of signal a state agriculture department could actually act on. And building a fully governed, lineage-tracked Delta Lake medallion pipeline under Unity Catalog in a single hackathon day is something the team hadn't done before at this speed.
What we learned
We learned that the hardest part of building for rural India isn't the AI — it's the data quality and the interface. APMC data is messier than any benchmark dataset. Prophet needs density to be reliable, and real-world mandi data is anything but. And every UX decision that seems obvious to a developer — type your query, scroll for results, read the chart — is a barrier for a farmer with low literacy on a 2G connection. Designing for voice-first and zero-typing changed how we thought about the entire product architecture, not just the frontend.
What's next for AgriSaathi
The immediate next step is voice-in, voice-out — replacing the text interface entirely with Whisper ASR so a farmer never needs to type. On the forecasting side, we want to move from Prophet to a Spark-native time-series model that can train across all crop-mandi combinations simultaneously, improving cold-start performance on sparse data. Beyond that, integrating Sentinel-2 satellite NDVI data would let a single leaf photo be cross-referenced with field-level crop health from space. Longer term, AgriSaathi's query log becomes a live epidemiological dataset for Indian agriculture — the outbreak detection layer, connected to Krishi Vigyan Kendras and state departments, could become an early warning system that operates at national scale.
Log in or sign up for Devpost to join the conversation.