Crop advisor lite

Pakistan has over 8 million smallholder farms of 1-12 acres, and most of the farmers running them have no affordable access to agronomic advice. Which crop to sow after wheat harvest, when the cotton sowing window closes, what an acre of mungbean costs in inputs, what to plant when canal water runs short - answers travel by word of mouth, often late and often wrong. One bad season can erase a smallholder's yearly margin. Crop Advisor Lite answers those questions with decision-ready guidance: named crops, sowing months, and rupee cost ranges. It understands questions in English or Urdu and always answers in English. Answers are grounded, not guessed: each question retrieves relevant passages (BM25 with crop-aware routing) from 40 Pakistani agricultural documents - government policy analyses, the Economic Survey, PCRWR research, the FAO crop calendar - and cites them inline. When the corpus doesn't cover a question, the app says so instead of inventing a citation, refuses to invent market prices, and deterministically refers pesticide-dosage questions to the local extension office - a code-level guarantee, not just a prompt instruction. Accuracy is measured, not claimed: the app scores 16/16 on a reproducible evaluation harness whose key facts come from the corpus documents, including honesty and safety probes (eval/run_eval.py, answers audited in eval/results.json). The stack: a Streamlit chat with streaming responses, running on gpt-oss-120b (OpenAI's open-weight model) via Fireworks AI - genuinely running on AMD-hosted inference, not just targeting it - packaged as a single Docker container with a public repo anyone can run. Next: Urdu-language responses and voice input, WhatsApp delivery, and a larger corpus of provincial extension guides. Open-weight models on AMD-hosted inference keep per-farmer cost near zero - the unit economics that could make good agronomic advice effectively free.