About

I build AI agent systems for work that people stake their reputation on.

Why vuca.dev

VUCA is the U.S. Army War College's name for an operating environment: volatile, uncertain, complex, ambiguous. It was coined for strategy after the Cold War. It also describes software development as an industry, architecture design as a practice, and applied AI in particular.

Models improve in depth and breadth of capabilities. Development requirements and best practices do not hold still. Agent output can be fluent and still be wrong. System design demands an elegant fusion of deterministic and probabilistic components.

These are exactly the kinds of environments I like to work in.

I'm a two-time technical co-founder, and a forward-deployed engineer by instinct. My method is the same in every domain: learn the work from the people doing it, then act as their technical partner to build software they rely on. Most recently that domain is financial audit, where I've grown accustomed to spending roughly a third of each day with the auditors who use our platform.

Every product I've built from enterprise to small businesses has involved discovery in the customer's domain, implementation against their environment and constraints, delivery measured by their outcomes and adoption, while bridging the technical and communication gaps throughout the process.

The agents I build operate where a wrong answer costs professional reputations. That constraint has driven my engineering discipline: holding accuracy steady over long contexts, graceful recovery against failure modes, making agent decisions inspectable, and guarding what deserves a human's attention.

Selected work — Punchcard

Backed by Y Combinator, General Catalyst, and Dell Technologies Capital

Financial Statement Review

Released as Punchcard's flagship agent. It uses distributed agents to run thousands of checks on financial statements inside live audits. I led it from discovery to release, redesigning the approach until it cleared a 95% accuracy bar. It became the fastest adopted product on our platform.

Read →

CoAudit

The platform's general-purpose agent. Auditors speak to it in their own vocabulary and it translates into the platform's. The harness uses progressive disclosure to route prompting and context, including a library of 80+ tools. It replaced a workflow-builder interface users found too abstract and cut time-to-output by 84%.

Creed

An autonomous engineering agent built to operate at a peer-level. It finds, triages, fixes, and merges its own code behind adversarial review and deterministic safety gates: around 30 merged pull requests a week, most with no human in the loop, at a revert rate under 1%.

Read →

GTM & Post-sales

Our largest customer, a 2,700-person audit firm, stalled in pilot because auditors were hesitant to trust LLM outputs. I built reasoning traces and a review surface that made every conclusion inspectable. They moved from pilot to an annual contract and reported cutting their evidence testing time in half.

Before Punchcard

FairQuote

Quoting and lead-generation software for local service businesses, which I co-founded and led engineering for. I trained a YOLOv5 model against mapping imagery to identify dying residential trees, generating 5,500+ qualified leads worth an estimated $8.2M, and replaced multi-day manual quoting with an instant, self-serve platform that processed $2M+ in quoted jobs.

See →

U.S. Air Force

Nuclear and missile operations officer. I led cross-functional teams responsible for up to 200 ICBMs across 23,500 square miles. Both operational missions as well as software, cryptography, and hardware modernization efforts.

Contact

Reach me on LinkedIn.