Senior AI Engineer

Led technical discovery with client executives and engineering leaders to identify automation opportunities, assess data readiness, define agent roadmaps, and mentor engineers through architecture and code reviews. Architected a reusable LangGraph agent layer with checkpointed state, typed tool contracts, resumable runs, and human approval gates for healthcare and compliance workflows involving FHIR APIs and systems of record.

Rebuilt RAG systems using section-aware parsing, hybrid BM25 and pgvector retrieval, cross-encoder reranking, and retrieval evaluations. Standardized guardrails with Pydantic-validated outputs, PHI redaction, prompt-injection defenses, and tool allowlists. Implemented release-gated evaluations in GitHub Actions, Langfuse and OpenTelemetry tracing, and model routing across OpenAI, Anthropic Claude, and AWS Bedrock, reducing LLM spend by 55%. Built voice agents, financial research workflows, Next.js reviewer consoles, Node.js MCP servers, and AWS ECS Fargate deployments using Terraform and blue-green releases.

Led a six-person engineering team across backend, frontend, and machine learning, overseeing hiring, sprint planning, design reviews, and technical direction for an AI search and assistant product from prototype to public launch. Owned the RAG pipeline, including query rewriting, web and Pinecone retrieval, PyTorch fine-tuned cross-encoder reranking, streamed cited answers, and an asynchronous FastAPI LLM gateway with provider fallback, circuit breakers, prompt caching, and cost tracking.

Built real-time Node.js and TypeScript services using SSE, WebSockets, Redis pub/sub, and connection draining. Developed a GraphQL gateway, multi-tenant ingestion with Celery and Kafka, and a Next.js/React platform using Tailwind CSS and TanStack Query. Optimized PostgreSQL through partitioning, indexes, and query rewrites, reducing thread-load latency from 4.2 seconds to 0.8 seconds. Owned EKS production operations with Docker, Helm, OAuth2, JWT, Datadog, offline evaluation gates, and comprehensive pytest, Jest, and Playwright testing.

Built a screenwriting assistant from its initial prototype in partnership with the founder and early user testers. Evolved the product from GPT-2 to GPT-3 and GPT-3.5, and fine-tuned GPT-Neo and GPT-J with Hugging Face Transformers and PyTorch. Improved screenplay-format accuracy from 87% to 92% using mixed precision and gradient checkpointing on constrained GPU resources.

Designed long-context memory using rolling act summaries and character state stored in PostgreSQL. Implemented generated-content safety using transformer classifiers, CLIP similarity checks, and YOLO detection, and added Tacotron2 character voices and Whisper dictation. Built the product stack with Next.js, React, TypeScript, FastAPI, Supabase, Node.js WebSockets, and Yjs collaborative editing. Ran model training on SageMaker, deployed Dockerized GPU inference, improved generation speed by 28%, and established GitHub Actions, Vercel previews, and Sentry monitoring.

Shipped customer-facing retail features using React and TypeScript on top of high-traffic Java services, partnering with product managers and designers from specification through A/B-tested release. Built Node.js and Express backend-for-frontend services that coordinated multiple Java dependencies with timeouts and partial-response fallbacks.

Engineered serverless APIs with AWS Lambda, API Gateway, DynamoDB, and CloudFormation, including single-table designs for peak-event traffic. Productionized scikit-learn and TensorFlow models with applied scientists, built near-real-time Python data pipelines using Kinesis and S3, and improved Prime Day readiness through caching, load testing, CI/CD hardening, and on-call participation.