Independent product · AI · Agriculture
NilamAI
A standalone retrieval-augmented agronomy SaaS that answers crop questions only from curated, source-weighted documents. NilamFlow is currently how users reach it.
- What I owned
- I designed and built it: the harvesting and extraction pipeline, the knowledge model (document, evidence, fact, card), the ask API, the admin interface, auth and API keys, deployments, the ingest tooling and the evaluation harness.
- What made it hard
- Sources arrive as PDFs, web pages, office files, scans, audio and video. A first four-agent extraction pipeline broke on real documents. I redesigned it as one LLM pass plus deterministic checks, and every fact must cite a source page. Embedding at ingest scale hit provider rate limits, and hosting moved several times before settling on Cloud Run.
- How it works
- Retrieval combines nearest-neighbour and keyword matches.
- Cards are ranked by score times source authority: scientific and government sources above commercial and farmer sources.
- Facts with unresolved contradictions are withheld.
- The model answers only from that context and cites card titles.
- Each response returns its sources with page-level evidence references, the facts it withheld, and confidence notes.
- Decision
- Fail closed. NilamAI answers only from curated knowledge and leaves out contested facts. The model is instructed to say when the context is insufficient and refer the farmer to an extension officer.
- Tradeoff
- Less coverage and fluency in exchange for traceability, which meant sustained work filling the knowledge base. The hard refusal fires only when retrieval returns no cards; otherwise it relies on the model following that instruction.
- What shipped
- A production deployment on Cloud Run, with a consumer API (service keys and rate limits) and an admin interface. Photo diagnosis is live in production, and access is currently limited to NilamFlow users. The admin Ask view shows the model, a confidence indicator, the cards used with their top facts, and the top excerpts with similarity. A preview mode runs retrieval without the model. It does not yet link through to source documents.
- By the numbers
- 5,400+ completed documents
- 57,000+ knowledge passages
- 280+ active source entries
- 14,600+ structured knowledge cards
- Grafting and budding collection: 290+ documents and 750+ passages, covering propagation, seasonal timing, nursery practices, tools, techniques and aftercare
- What improved
- Every answer carries a trail back to the cards, facts and source pages it was built from, so a person can check it instead of trusting it.