Building the AI layer for hazardous waste compliance: a multi-service platform that classifies waste, extracts and validates regulatory documents, and automates vendor portals, serving 200+ customers including a production agent shipped to a $3B enterprise.

• Shipped an AI waste-classification agent for Veolia end-to-end, fine-tuning Gemini on Vertex AI and deploying on AWS, generating 20K+ structured waste profiles per month
• Designed a human-in-the-loop manifest workflow that cut document processing time 85%, combining PDF extraction, AI validation, and EPA e-Manifest automation
• Built Profile Assist, a LangGraph agent running LLM inference on Baseten, with LangSmith tracing the LLM internals (tests assert per-run token and cost tracking against Baseten pricing metadata) and Sentry covering API and backend failures
• Built the QA/QC layer for the AI pipeline: structured-output contracts, deterministic regulatory rules, confidence-based routing to human review, and golden eval sets built with in-house waste-compliance experts, re-run on every model and prompt change
• Built the WASTELINQ MCP server, giving Claude and ChatGPT read-only access to customer reporting data: FastMCP on AWS Bedrock AgentCore, Cognito JWT auth with per-broker isolation, fed by a nightly Celery snapshot job with task-failure and freshness alarms
• Designed that server's context layer as tiered progressive disclosure: 21 report templates and ~700 fields served as list, then describe, then per-field detail, plus an on-demand agent-facing wiki, so a model loads only the context a question needs
• Built a 7-microservice AI platform (Lambda, Fargate, Terraform) plus a portal-automation product spanning 4 disposal-facility vendors

Founded and built an enterprise AI evaluation platform solo: an LLM-as-judge system that scores manager coaching sessions against a leadership competency model, backed by a multi-tenant Supabase architecture and a quantitative ROI engine.

• Designed an LLM-as-judge evaluation engine using the OpenAI Realtime API for live voice coaching sessions, with deterministic scoring and a full assessment-state lifecycle
• Architected a multi-tenant SaaS on Supabase (Postgres, row-level security, 18+ Deno edge functions) spanning coaching, billing, and reporting
• Shipped a full production stack solo, with a test suite that runs through live voice sessions as well as the web app

Led technical strategy for a multi-modal generative video platform as CTO, taking VisualLabs from a Colab prototype and Techstars '23 pre-seed funding to a production system handling 100k+ monthly renders.

• Took the product from a Colab notebook to a production video platform on AWS, re-tuning the diffusion pipeline and moving orchestration onto Step Functions and Lambda to cut render time 65%
• Provisioned burst GPU capacity instead of always-on clusters to keep per-render cost viable at 100k+ renders per month
• Shipped per-customer personalization on two separate LoRA stacks, one fine-tuned on individual people for identity and one on a visual aesthetic for style, composed with ControlNet temporal conditioning. It worked technically, but face fidelity was not where the market needed it yet
• Ran the entire engineering function as the only full-time engineer, directing one contractor and two interns