A selection of recent work. The Gauntlet systems below were built end to end and most ran live; the partnerships are teams I embedded with through Dev Forward.
A recursive software factory where every unit of work is a typed goal. One recursive operation drives the machine: decide, evaluate through a gate plus an LLM judge, repair, escalate, report upward. All state is a projection of an append-only event log.
A source-cited AI agent embedded in the patient chart that writes a 90-second pre-visit briefing and answers grounded follow-ups. Every claim is verified against a typed snapshot before it reaches the UI; uncited claims are dropped at the gate.
A continuously-running adversarial multi-agent platform that probes the Clinical Co-Pilot across six attack categories. Four independent agents coordinate over a typed message bus; it found and documented four confirmed vulnerabilities running live.
A non-custodial, same-day binary options market on Solana for MAG7 stocks. A custom on-chain order book, PDA-owned vault and mint authority so no human ever holds user funds, and Pyth pull-oracle settlement.
A voice-first assistant for aircraft-maintenance workers. Its defining thesis is a grounding gate: any factual claim not backed by a tool result that turn is refused aloud rather than guessed, down to per-field numeric verification.
A fully self-hosted, low-latency multi-channel push-to-talk system with no managed RTC platform. A Go/Pion SFU owned end to end: signaling, STUN/TURN, media forwarding, observability, deploy. Splits signaling from media to scale horizontally.
A real-time mastery interface teaching Boolean logic across three irreducible representations at once. Mastery is gated by Bayesian Knowledge Tracing plus a response-time band, a held-out transfer item, and an LLM-judged explain-back.
A detect, obfuscate, call-LLM, restore pipeline that lets users query external LLMs over sensitive documents without raw PII or PHI crossing the trust boundary. A per-session vault preserves co-reference while the model sees only pseudonyms.
A precision roof-measurement system fusing satellite imagery, public LiDAR, parcel data, and guided phone capture into accurate roof geometry. The hard problem: fusing heterogeneous geospatial sources with accuracy guarantees under obstruction.
An AI tool that turns litigation source documents plus a firm template into a format-faithful demand-letter draft where every factual claim is cited to the record and every gap is flagged. De-identification runs before anything reaches the LLM.
A LangGraph multi-agent fitness-coaching system with streaming, multi-turn memory, and domain constraints like injury avoidance and bilateral muscle pairing. A compact demonstration of disciplined orchestration, backed by 258 passing tests.
A reusable AI code and security reviewer that drops into any GitLab merge-request pipeline with a one-line include. A cheap model triages every file; high-risk or security-sensitive files escalate to an expensive model for adversarial review.
A conversational math tutor teaching fraction equivalence to nine-year-olds, running on iPad. A voiced, scripted tutor syncs with an interactive chocolate-bar manipulative. The LLM is kept genuinely optional behind a hard determinism boundary.
A persistent, self-curating memory system a coding agent installs as a Claude Code plugin, plus the eval framework that proves it helps. Built with a team of four; I own the harness architecture, the core run loop, data models, CLI, and plugin MVP.
As RentButter accelerated client acquisition and product velocity, I embedded with the team to ensure the platform could scale responsibly. We stress-tested APIs, modernized development practices, and aligned technical architecture with the long-term roadmap.
Improved system reliability, stronger unit economics, and infrastructure designed to support continued growth.
Camino had launched but was still refining product-market fit. I joined as a fractional technical partner, contributing hands-on engineering, strategic product input, and ownership of technical go-to-market initiatives including SEO and AEO.
A market-ready product, early customer traction, and a team focused on scaling the business, not managing technical overhead.