Agronomy-specialised Large Language Model for Vietnamese Crop Advisory and Soil-Sensor Context Understanding 

Enfarm FAMOS AI is a Vietnamese-language agronomy assistant built for durian smallholder farmers in Vietnam's Central Highlands, developed as an RMIT capstone in partnership with Enfarm, an IoT soil-sensor company. Farmers interact through text, voice, or leaf photos, receiving answers grounded in their own farm's live sensor data and a curated agronomy knowledge base rather than free-form LLM generation.

The architecture separates real-time FAMOS sensor context, a Qdrant-backed agronomy knowledge base, and deterministic YAML skills that answer common questions such as farm status, irrigation, and disease care without invoking the LLM at all. The model is reserved for open-ended queries and is always constrained to retrieved sources. The vision module avoids asking the model to name diseases directly. Instead, it generates a structured Vietnamese description of leaf symptoms, then matches it against an eight-class symptom library through retrieval, falling back to "unknown" rather than guessing when confidence is low. Voice input runs through Google Cloud Speech-to-Text in Vietnamese, phrase-boosted for durian vocabulary, producing a transcript the farmer confirms before it enters the same chat pipeline as typed text.

Built on FastAPI, Streamlit, Qdrant, and Postgres, and fully Dockerized, the system is designed to keep answers safe, explainable, and grounded in real farm data.