I build production-grade AI systems that solve problems that actually matter — under real operational and legal pressure, for real clients.
Not proofs of concept. Not demos. Production systems and open-source tooling running under real operational and legal pressure.
Converts fragmented HOA and condominium governing documents into citation-grounded, meeting-ready answers. Three-stage OCR pipeline, hierarchical document weighting, and a proprietary confidence scoring algorithm that shows exactly how every answer was derived — so a board member who doesn't trust AI can see the work.
SHA-256 chain-of-custody ingestion across a 110,000+ document corpus spanning PDFs, spreadsheets, Outlook email, and SMS corpora. Context-window-aware subagent scaffolding. AI-orchestrated case analysis mapped to specific legal complaint counts. In active production use by a law firm in civil litigation defense.
An append-only agent memory store grows forever, costing more to read the longer you work. A read-only sidecar folds raw observations into one evolving digest per project category — rewriting understanding in place instead of stacking notes. 137× compression on the heaviest project; flat read cost regardless of history depth.
A compounding stack of workflow automations — format-agnostic invoice ingestion with direct API integration, license plate recognition for parking enforcement, governing document Q&A with hallucination guardrails, and a voice-matched correspondence assistant trained on 40,000+ personal documents — that collectively reduced operating overhead by 41%.
The scoring engine behind BoardPath, extracted into a standalone npm package.
Turns the signals a RAG pipeline already produces into a 0–100 scorecard
where every point has a reason attached — and returns an answer /
review / abstain action, not just a number. Eight
dimensions, 412 tests, dual ESM/CJS, zero runtime dependencies. Apache-2.0.
I build systems that solve real problems under real pressure. Not prototypes. Not demos for the sake of demos. The things I've built have been used by law firms in active litigation and are weeks away from public product launch. Real pressure produces real design decisions.
My approach starts with the problem, not the technology. I spend more time understanding what's actually broken — and why — than I do selecting tools. The architecture follows from the problem definition. When you start with the tool, you end up bending the problem to fit it. That's how you get AI systems that look impressive in demos and fail in production.
I design for reliability, transparency, and human oversight. AI systems that don't tell you when they're uncertain, don't cite their sources, and don't keep humans in the loop where it matters aren't useful — they're liability. Every system I build reflects that conviction.
We're at an inflection point for this technology that is comparable in scope to the Industrial Revolution. The builders who matter in this moment aren't the ones who can predict exactly where it goes. They're the ones who understand the underlying mechanics deeply, build things that actually work, and keep building. That's the work.
Short-form posts on specific architectural choices from real builds. How I think, not just what I shipped.
Open to full-time, contract, and fractional engagements across AI systems architecture, agentic workflow design, and document intelligence. Remote-first. Available for hybrid with reasonable travel.