Software engineer. Python primary, JavaScript secondary. These days I point that at AI systems: LangGraph agent orchestration, retrieval pipelines, and multi-provider LLM routing, plus the full-stack product around them.
Most of what I get called for is the same problem. A team bought or built an AI feature, it demoed beautifully, and it does not survive contact with real users. That is rarely a model problem. It is state, retries, fallback, and deciding what the system is allowed to do when it is unsure.
What I work on
- Agent orchestration and state management with LangGraph, including real human-in-the-loop interrupts
- Retrieval over public records, where the data decides and the model explains
- Multi-provider LLM routing and graceful degradation
- The boring parts: tests, citations on every number, claims that match the shipped defaults
Shipped and running
- MCGrantz: grant winnability scored against USASpending award history and IRS 990-PF filings. FastAPI, Postgres, Claude. Live
- MCStanding: data-broker compliance across 11 regimes with no backend, so regulated data never leaves the browser. React, 304 tests. Live
- HumxnMed: patient health briefings behind a LangGraph guardrail that routes crisis input out before generation. Live
- MCProof: Section 508 and security scanning that drafts the VPAT/ACR. Live
Open source. Small, complete, tested tools built on patterns I run in production:
- failclosed-guardrail: a LangGraph safety guardrail that fails closed. Crisis input exits before generation, the human-in-the-loop step is a real interrupt, and CI fails if any crisis case leaves the safety routes. 130 tests.
- llm-failover: multi-provider LLM routing with retries, per-provider circuit breakers and a deadline. Falling back never hides a revoked key. Zero required dependencies, 100 tests.
- award-odds: a CLI that shows who actually won a federal grant program, using public USASpending data. Transparent rules, no LLM in the scoring path.
The product source is private because it is commercial. The design reasoning is public in ai-engineering-portfolio. Earlier experiments: emosafe-ai, ai-failure-analysis.
Where Founder and AI and software engineer at Millennials Creatives.
Also Executive Director of Finesse Our Minds, a youth mental health nonprofit offering free peer support from people who get it. Years of that work taught me the thing I design around: detection is the easy part. What happens in the moments after the signal fires is where systems fail.