SATQUERY AI
SOLE AUTHOR
An agentic vision-language assistant for satellite imagery whose every number cites its evidence.
81.8%
ACCURACY · 100% CITED
- answer accuracy, zero-shot to fine-tuned
- 26.4 → 81.8%
- accuracy on 500 held-out samples
- 80.7%
- citation precision
- 100%
- uncited numbers
- 0%
COMMITS18 / 18
THE ACCOUNT · summary
It started as a Smart India Hackathon entry for problem statement 26167 (ISRO / Space Applications Centre) and is now my own project. Inputs pass 11 compatibility checks, a fixed policy table runs specialist tools in order, and every answer ships with a versioned audit trace.
CLUES FOUND · what I built
- A QLoRA fine-tune of Qwen3-VL-8B on 53,098 samples from BigEarthNet-v2, VRSBench, RSVQA-HR, CDVQA and Evidence QA.
- Siamese change detection for before-and-after imagery, and fusion of optical and radar (SAR) inputs.
- Fully local serving on one 24 GB GPU, plus a 5 GB quantised build for an offline 8 GB laptop.
- CI with ruff, mypy --strict, pytest and vitest, and a contract check that keeps API and front-end types in sync.
FURTHER EXHIBITS · more screens
TOOLS RECOVERED · stack
- Python
- FastAPI
- rasterio
- GeoPandas
- PyTorch
- Transformers
- PEFT
- llama.cpp
- React 19
- Tailwind v4
- Docker