At Zoop.One, I owned training, inference, deployment, monitoring, and rollout for a production face-liveness system serving 5K–6K requests a day. I replaced a third-party liveness API with in-house GPU inference, reducing annualized serving cost by over 90%.
I also built motor-claims document classification and OCR workflows, with a field-level accuracy framework that scored 759 documents at 96% exact-match accuracy. By improving claims intake and implementing semantic document retrieval, I helped cut manual review time by 35%.
At Qualcomm, I integrated torch.compile into neural network quantization libraries, achieving a 20% runtime speedup and 15% GPU memory reduction. My project work includes a knowledge-distillation pipeline for 500K+ noisy production strings and deep-learning research on channel estimation.

