What I'm focused on

Shipping

• Looking for the next role. Most recently I was the sole engineer on an enterprise healthcare deployment at Kinetic Systems in SF, which wrapped in May.
• Eight systems across four disciplines, built this month and shipped to GitHub: Raft consensus in C++20, a three-broker task queue, a GraphRAG reasoning engine, a sandboxed code-repair agent, an inference gateway, a streaming orchestrator, and two forward-deployed data systems.
• Closing the measurement gap on those repos. Four carry real benchmark numbers; the rest describe architecture until their measurement protocol has actually been run. A bracketed placeholder metric is worse than no metric, so they stay unnumbered until the number exists.

Reading

• Natural Language Autoencoders— Anthropic's paper on steering LLM representations, and the clearest read yet on the gap between what a model says and what it internally represents.
• The Mom Testby Rob Fitzpatrick — re-reading it with the lens of "how do I validate EEAAO without building the wrong thing first."
• Anthropic's Constitutional AI papers and the broader alignment literature — keeping current with the field.

Thinking about

• How much of "semantic" caching is actually semantic. Building an LLM gateway, I swept the similarity threshold against adversarial near-misses — Australia vs. Austria, enable vs. disable — and found near-misses sitcloser in embedding space than genuine paraphrases. No threshold makes that cache both useful and safe. Negative results are still results.
• Whether on-device AI (Gemma 4, Phi-4) is underrated for personal productivity apps — privacy by default, zero latency, no API costs.
• The evaluation gap in healthcare AI: most teams ship models without rigorous eval pipelines. HealthAdminBench is one attempt to close this. There are a dozen more benchmarks that need to exist.
• What "founding engineer" actually means at an applied AI startup — and which companies are building real infrastructure vs. GPT wrappers with good PR.

Looking for

Founding-engineer or ML engineering roles at applied AI startups, Series A–C, in SF or remote. I care most about the problem domain and the quality of the team. Healthcare AI is where I have depth, but I'm not limited to it — anywhere that rigorous evaluation and real-world deployment are first-class concerns.

Not doing

• Taking on freelance work — fully committed to the above
• Speaking at conferences (not this year)
• Social media that isn't directly build-in-public related