AI is powerful but unpredictable. We constrain it with validation layers, schema enforcement, and human review at critical junctures. The result: systems you can trust in production.
AI outputs are proposals, not facts. Every output is validated against strict schema and logic constraints before it enters your workflow.
We design workflows where humans review critical decisions, optimizing for high-leverage intervention rather than full automation.
The real world is messy. Our systems ingest unstructured data—texts, emails, PDFs—and structure it for downstream systems.
Open-source MCP server that gives Claude native access to six macOS applications. Three purpose-built backends—EventKit, JXA, and SQLite—each optimized per app. 95% test coverage across 25,000 lines of TypeScript and Swift.
Transform scattered incident records into OSHA-compliant forms. Handles messy inputs—texts, emails, spreadsheets—and generates Form 300A + ITA files.
How to build AI systems that produce the same correct output for the same input. Testing strategies, validation approaches, and patterns for production-ready AI.
How contract-driven development prevents scope creep, enables binary verification, and ships compliant systems faster through upfront clarity on success criteria.
How a filesystem-based ledger pattern enables multi-agent AI coordination with deterministic state, no database complexity, and perfect auditability.
Code over consensus. Measure before you optimize. Read before you edit.
I build AI systems that bring engineering rigor to probabilistic models. My approach: validation layers, human-in-the-loop workflows, and deterministic constraints that make AI agents reliable enough for production.
From multi-agent orchestration to MCP server development, I specialize in translating complex rules into workflows that actually work. Calm, practical, precise.
About My WorkLooking for help with AI systems architecture, multi-agent orchestration, or deterministic design? Get in touch.
Get in Touch