I've built production AI systems at Hindustan Times, BHEL, and the National Informatics Centre, from fine-tuned LLM inference endpoints that reduced latency by 40% to computer vision workflows processing 10,000+ land-record images at 92% classification accuracy.
I implemented GRPO from scratch to train a 1.5B-parameter reasoning model, achieving a 34% relative gain on GSM8K and MATH. I also engineered a multi-agent software repair system evaluated across the full SWE-bench Verified benchmark.
I work across PyTorch, Hugging Face, vLLM, LangGraph, Docker, Kubernetes, RAG, and MLOps, with hands-on experience taking models from training through scalable deployment.
