I'm a Computer Science and Data Science senior at Knox College. I build full-stack and systems software, mostly products that put machine learning in front of real users. On the side I do geospatial ML research with satellite and drone imagery, which keeps me sharp on the modeling side of things.
- 🏆 1st place at HackIllinois 2026, Best Use of Actian VectorAI
- 📦 Published bhasha-js on npm, an internationalization SDK for South Asian languages
- 🎓 Break Through Tech AI Fellow at Cornell Tech
- 🛰️ Five research papers (three under review) on satellite and drone based environmental monitoring
- 🌐 More at thesantoshpant.github.io
| Project | What it is | |
|---|---|---|
| Vigilant AI | Turns raw CCTV footage into something you can actually search. Runs YOLO, CLIP, and a vision-language model in parallel on a serverless A100. Won HackIllinois 2026. | repo |
| CropScan | Crop-disease diagnosis for smallholder farmers. A DINOv2 + EfficientNetV2 ensemble that backs off and says "not sure" when the models disagree, so it doesn't hand out confident wrong answers (FastAPI, React, Docker). | cropscan.tech |
| OMS | A from-scratch in-memory crypto order-matching engine in Go, around 4.2M orders/sec, with FIX 4.4, event sourcing, and pre-trade risk checks. | theoms.vercel.app |
| bhasha-js | A drop-in i18n SDK that serves AI-generated translations with no per-language setup work. | npm |
| Yaar | A full-stack AI study-abroad counselor with a human-approval step and spend controls built in. | okyaar.vercel.app |
| Terai Heat Forecaster | Early-season heat-stress forecasting across 21 rice districts in Nepal (Google Earth Engine, PyTorch). | live demo |
Languages: Python, Go, TypeScript Web: React, Node, FastAPI, PostgreSQL Infra: Docker, AWS ML: PyTorch, Google Earth Engine, Sentinel / MODIS / Landsat
A smaller part of what I do, but it feeds the rest. I care about honest baselines and models calibrated enough to actually make a decision with.
- Heat-stress forecasting over the Terai, where last year's values turn out to be a tough baseline that fancier models often can't beat.
- Geospatial foundation models: when a frozen encoder is just as good as full fine-tuning, and how proxy labels can quietly inflate benchmark numbers.
- Conformal prediction for district-level crop-yield loss, with finite-sample coverage guarantees.
- Predicting thermal heat maps straight from ordinary RGB drone imagery, no thermal camera required.
Three are under review and two are in preparation. Venue names are off while things are under blind review, but happy to share details if you ask.
- LinkedIn: the-santosh-pant
- Portfolio: thesantoshpant.github.io
- Email: pantsantosh23@gmail.com