At 4Sight AI, I shipped a RAG system that audits POCSO charge-sheets against more than 1,000 legal documents and provides cited evidence. It’s live with AP Police in three pilot districts, and its scorecards and evidence-gap reports take under five minutes per case.
I also built real-time face recognition for a 2,000-camera police CCTV network, reaching 93.3% true-accept at a 1-in-1,000 false-match rate. For traffic enforcement, I trained YOLO26 on more than 50,000 camera images to flag no-helmet violations.
At GGS Information Services, I shipped 3D mesh compression for more than 15 enterprise clients and built a GAN-based 2D-to-3D CAD pipeline. Quantization and CUDA kernels helped reduce inference latency by 30%.
In open source, I created indictok, tokenizers for 22 Indian languages and 12 scripts, and reproduced GSPO post-training for Qwen. My work also includes a first-prize AI4 Andhra Police Hackathon project and a paper on identity-preserving image synthesis.
