system init / profile loaded
Models that hold up.
Researching deep learning for computer vision, with a focus on egocentric vision. Building models that survive contact with real data — from 99.7% AI-audio detection to MVSS-Net Lite forgery detection. Finished breaking things at CDAC. Still fixing things that matter.
system logs / experience
Four stops.
Document forgery, audio classification, campus communities, and one salary model that had to survive held-out data.
Built DocForge, a FastAPI document-forgery detection application from upload to manipulation classification. Integrated MVSS-Net Lite for scanned-document inference.
Delivered monthly DSA Saga sessions for a 250+ member club, sustaining 40–50 attendees each month.
Drove campus-wide Adobe tool adoption through technical workshops for 100+ students.
Engineered an XGBoost + scikit-learn salary prediction model achieving 89% R² on held-out test data.
telemetry / selected artifacts
Four artifacts.
Hover a record. Inspect the signal. The interesting part is usually hiding in the moving pieces.
human systems / community
The work is not only the model.
Sometimes it is the room around the model — who gets invited in, who gets unblocked, and who stops pretending DSA makes sense by itself.
50+ member community, 30-person core tech team, ML/web/DSA curriculum, workshops, hackathons, and speaker sessions. The job was mostly making the next good thing easier for someone else.
For underclassmen. Show up, solve the problem, and stop pretending DSA makes sense by itself.
Stanford CS231n + Karpathy Zero-to-Hero — micrograd and makemore rebuilt from memory. Research direction: egocentric vision.
offline signal / people around the work still count as infrastructure
signals from the field
A few places, communities, and teams that made the work more real.
Adobe
CDAC
AARVAK
HackerRank
LaunchEd