← All work

Bhoonidhi

Nov 2025 - Dec 2025

Satellite imagery you can ask questions about in plain English. National top five at an ISRO hackathon.

The problem

Working with satellite data normally means QGIS, coordinate systems, spectral bands and a machine that can open a 10,000 pixel wide image. Most people with real questions about land never get past that wall.

What I built

We turned it into a map and a question box. Upload Sentinel-2 archives, click around, ask things in plain English.

I led the team of six and worked on the LoRA fine-tuning, the raster pipeline, the routing layer and the frontend. The fine-tuning was the fun part: we taught an open multimodal model to actually read satellite imagery.

How it went

71.26% VQA accuracy and 0.415 BLEU-1, beating several open-source baselines that had far more compute than we did. We finished top five in the country at a national ISRO hackathon.

Numbers

VQA accuracy
71.26%
BLEU-1
0.415
National finish
Top 5

Python · PyTorch · LoRA · FastAPI · React