CropAdvisor

THANJAVUR CROP ADVISOR: a Tamil voice crop-decision helper The problem In the Cauvery delta, a farmer choosing the next crop has to weigh uncertain prices, rainfall and water supply, often with patchy internet and without time to work through spreadsheets. Most digital tools assume the farmer can read and type in English or navigate menus. What it does The farmer tells the app, by voice in Tamil, the season, the number of acres and the irrigation water they can count on. The app ranks paddy, black gram, groundnut, sesame and maize, showing for each crop a profit range from a poor season to a good one, a risk label and any water shortfall. The farmer can then ask follow-up questions in Tamil, such as how the advice changes if rain is late, and hears the answer spoken in Tamil. How it works - Muse Voice Transcribe turns the farmer's spoken Tamil into text. - Muse Spark understands the question, calls our decision engine through tool calling, and explains the result in Tamil. - A deterministic engine (SQLite plus calculators) computes every number. The model never does arithmetic, so the figures stay reliable and checkable. - Ranges replace single numbers, because prices and rain are uncertain. - If the internet or the model is unavailable, the app falls back to the on-device engine and a short templated Tamil summary read aloud by a local voice. Voice input and model answers need a connection. - The API key stays on the local server and is never sent to the browser. Data and honesty The demo uses simulated, clearly labelled sample data, and the app displays a notice saying so. The real-data pipeline is built and tested: Agmarknet daily prices convert to monthly medians, and crop costs (from Tamil Nadu cost of cultivation tables), yields and rainfall load from CSV files. Limitations and next steps Tamil wording has not had a full native-speaker review, and the app has not been tested with farmers. Next steps are loading real data for more districts, a Tamil-speaker review, and testing with farmers.