Stuff I've actually built. Most of it links to code or a live demo.
Comparing LLMs (Qwen2.5-VL) vs CNNs for transcribing historical documents. Python layout engine for multi-column reconstruction.
Altair dashboard with 114,000+ Spotify tracks. "Mood Maps" (Valence vs Energy), brush-linked selection, multi-genre filtering.
Research pitch for fencing performance tech. Motion capture, real-time feedback, market sizing.
Deep learning model hitting 87% accuracy on heart murmur classification from stethoscope audio. Runs on a digital stethoscope for cheap cardiac screening.
Legal guidance platform I co-founded. Case tracking, AI summaries, lawyer matching — live and used daily.
Combining ARIMA, Echo State Networks, and LSTM for time series prediction. Tested on Brent Oil, Gold, and Sunspot datasets.
Lip-sync pipeline doing 95% accuracy at 30 FPS, supports 4K. Built this for video dubbing at GoWarm.
Classifies 93 cetacean species from underwater audio using CNNs, CRNNs, and SoundNet. IEEE published + Indian patent filed.
Real-time GPS tracking for public buses. Built so students could actually figure out when their bus was coming.
Robot that talks back. TTS/STT with LLM integration so it can hold a conversation.
EM spectrum analysis payload for a 1U CubeSat. Redundant systems, weather-resistant build.
Dance floor that generates electricity. Air-tight syringes + piezoelectric sensors under the tiles.