40 tests on a real database
FieldProof
Field photos in, verifiable impact reports out
- Year
- 2026
- Role
- Owned backend, AI pipeline, demo data & deployment
- Context
- Code Cubicle 6.0 · team of 4
- Stack
- FastAPI, Python, PostgreSQL, pgvector, CLIP, LiteLLM, Next.js
Overview
A Python worker tags uploaded photos with CLIP, embeds them for pgvector search, assigns sites from EXIF GPS and estimates before/after green cover. AI-written reports that invent a number are rejected.
How it works
- Postgres-backed job queue feeding a worker that runs CLIP tagging and embeddings via ONNX.
- Semantic photo search on pgvector; sites assigned automatically from EXIF GPS.
- LLM output containing a literal number is rejected, falling back to the next model, then to a template.
- 40 pytest tests run against a real database, not mocks.
The live demo serves a saved snapshot of real data while the backend is offline.
Next project
DataPilot