Hi, I'm Chirudeva! I can summarize my experience and how I got here in just a few sentences:
- Was lured into going down the rigourous and "prestegious" academic path that honestly didn't interest me one bit in the start.
- Started loving the process a year back because I enjoy solving problems that bother me PERSONALLY, and my fundamental argument is that if I can solve problems for myself, why not optimize the process and do the same for others!
- Finally, living and learning in the era of Artificial Intelligence has taught me one thing - In a world full of frontier models, you most certainly can build cool things by outsourcing intelligence, but FORCING myself to take the extra mile and learn while building cool things is what makes this process exponentially better.
Click a project to open it up.
Body2Health · body measurements from two silhouette photos, plus the health indices you can derive from them
It works from silhouettes rather than identifiable photos, which is the privacy-conscious part. YOLO and SAM2 handle preprocessing, a dual-view model does the estimation, and an SMPL-X reliability gate rejects weak predictions instead of quietly reporting a number anyway.
YOLO SAM2 SMPL-X PyTorch Computer vision
Pre-generation hallucination detection · catching RAG hallucinations before the model emits them
A research pipeline that reads hallucination risk out of transformer hidden-state dynamics, before any hallucinated token is generated. Five internal drift signals, causal activation patching to check the signals are actually causal, and evaluation across domains.
Transformers RAG Activation patching Interpretability
Road accident severity prediction · severity models and crash-hotspot maps over NSW and Chicago data
An end-to-end ML and geospatial pipeline. DBSCAN finds the hotspots, SHAP explains what the model is keying on, and Plotly, Folium and Dash carry the results into something you can actually click around in.
DBSCAN SHAP Plotly Folium Dash
The stack
The snake eats my commit squares once a day. Here it is underwater.
Everything on this page is generated by .github/workflows/snake.yml and committed to the output branch: the snake by Platane/snk, the rest by tools/build_assets.py. No third-party image service is contacted when you load this page.
Hevy workout graph, weekly minutes over the last 16 weeks.
I am open to AI engineering roles and to collaborating on applied ML, trustworthy AI, health tech, and fitness tech. If you have a hard, practical problem, send it over.
Email me or connect with me on LinkedIn.
