Engineering leader building agent systems, applied AI for healthcare, and developer tools.
I'm a founding engineer at Architect Health and building a stealth startup. I like staying close to the build.
My work sits where autonomous agents, applied AI for healthcare, infrastructure, and product judgment meet. This site is a working notebook for what I'm building and what those systems are teaching me.
Selected work
Agent teams
A workspace where specialized AI agents plan, code, review, and resume work with scoped context and human oversight.
What it taught me: Agents get more reliable when they have structure: clear roles, budgets, worktrees, and review loops.
Read moreBehavioral systems
A set of behavioral guardrails for how agents build, verify, review, ship, and operate inside engineering workflows.
What it taught me: Reliability is an architecture problem, not a prompting trick.
Read moreInfrastructure
Local inference and private workflows for systems that should keep working without sending core data to hosted APIs.
What it taught me: The fastest way to keep control is to avoid renting the capability that matters.
Read moreResearch tooling
A plugin-based way to give agents the data they need without handing them one undifferentiated context firehose.
What it taught me: Context-scoping beats context-stuffing.
Read moreDeveloper feedback
Screenshot tooling that lets agents inspect the interfaces they build and catch visual issues tests miss.
What it taught me: If an agent can see what it built, visual feedback becomes part of the development loop.
Read moreWhat I am learning
Some come from building agent systems. Some come from healthcare and infrastructure work. The pattern is that real products make the hard questions harder to ignore.
See longer notesThe same workflow can behave differently under a new model even when the harness, tools, and prompt are unchanged.
Long-running work gets better when the system starts with a small index and pulls deeper context only when the task needs it.
For systems that summarize, recommend, or act, I trust layered evaluation more than a single pass/fail gate.
When agents do more of the mechanical work, the harder job becomes judgment: deciding what good looks like and keeping the loop honest.
Latest notes
Hosted LLM APIs are great until the vendor changes the rules. Here is why I run inference locally with Ollama, what it costs, and what you gain when the data never leaves your machine.
Prompt engineering gets you 60% of the way. Skill frameworks get you the rest. Here is how 21 structured behavioral guardrails turned inconsistent AI agents into reliable engineering teammates.
The role of software engineer is quietly being rewritten. When AI agents do the mechanical work, your job becomes orchestration, judgment, and design.