Multilingual crop advisory with documented language gaps

The Deccan Agri Observatory sample describes a voice-and-text advisory bot that recommends sowing windows and pest responses. Early logs showed confident answers in a majority language and hedged or missing answers in two minority languages spoken by a large share of smallholders.

Rather than hide the gap, the team shipped a public coverage map, a “do not use for pesticide dosage” warning, and a human agronomist callback for low-confidence replies. Plot-size bias was tested by holding soil and weather constant and varying landholding in the evaluation set.

The profile is useful for agriculture ministries comparing vendor claims: the innovation is not the chatbot itself, but the decision to publish where the model should not be trusted.