Led the development of COTA, a production LLM application for 120K+ customer support tickets monthly using GPT-4 with a custom RAG stack (Pinecone, OpenAI embeddings, BM25 hybrid retrieval). Implemented an end-to-end LLMOps pipeline with prompt versioning, A/B testing (GPT-4 vs Claude vs fine-tuned), evaluation (ROUGE/BERTScore/LLM-as-judge), and guardrails for PII detection and hallucination mitigation. Established production monitoring for token usage, latency, cost, and quality; deployed multiple models and managed routing to balance performance and cost. Achieved 10% faster resolution times and delivered 50M+ daily predictions across Uber's systems.

AI Solutions Engineer at Viz.ai

February 1, 2018 - November 1, 2021

Developed CNN-based stroke detection models (ResNet-50, EfficientNet-B3) for Large Vessel Occlusion detection from CT/CTA, achieving 94% sensitivity with 35% fewer false positives across 50K+ studies. Built real-time inference pipelines (<30s end-to-end) with TensorFlow Serving, Triton Inference Server, and Kubernetes, supporting 2–16 pod autoscaling and 99.7% uptime. Designed distributed training pipelines (TF, Horovod) on multi-GPU clusters, reducing training time from 48h to 6h for 200K+ images. Integrated models with PACS across 15+ facilities, enabling FDA-cleared AI solutions that saved radiologist time and reduced door-to-treatment times.