Now
- Leading a team at a network security company, responsible for security monitoring products from architecture through deployment and production operations.
- Pushing Autograph from limited distribution to a public 1.0, iterating on changelog-tracked feedback from real writing use.
- Fixing inference-core issues in oMLX, an open-source LLM inference server for Apple Silicon. As of July 27, 2026, two pull requests remain open and unmerged; GitHub is authoritative for their live status.
Selected work
- Fonos: Local speech-to-text, plus reusable processing for whatever text you have selected — all of it inside the app you are already working in.
- WindowPilot: A visual window navigator for people who keep more than one idea open at a time.
- Autograph: A local long-form writing tool that turns the heading structure of one plain Markdown file into a navigable, rearrangeable, compilable tree.
These are the three personal products currently approved for public presentation. Their records stay within facts supported by source, release material, or real interfaces, and clearly identify what readers cannot independently inspect when source or distribution remains private. Feature count is not presented as market impact.
Turning points
Early 2026 — present
Independent product builder
Began building independent products in early 2026. The public release records documented here for Fonos and WindowPilot begin in 2026; Autograph remains a limited-distribution pre-1.0 build. I also began contributing to oMLX. It was the first time I handed my own work to strangers — and the first time AI let my execution keep pace with my judgment.
2018 – 2019
Internal transfer to the U.S. (Staff → Principal Engineer)
Moved my family to California — the largest decision I have ever run through my own risk calculus. Since then, “controllable” stopped being a stated preference and became the first computation in every major choice.
2015
From network equipment to cloud platforms
Left five years of protocol-stack work on network equipment for cloud platforms. My concern shifted from the correctness of one protocol layer to how a whole system behaves under real traffic.
How I work
- I define the problem, write down the approach, and fix the acceptance criteria before I build — structure and abstraction come before code; the cost is a slower start, and low tolerance for figure-it-out-as-you-go settings.
- I converge problems into shapes with a definite outcome and a way to verify it — that is where I do my best work and enjoy it most; the cost is that truly open-ended problems make me anxious before they make me curious.
- Product pages state only what the source code and the real interface can jointly confirm; the cost is marketing that always reads conservatively.
- I hand execution to AI agents at scale, but never architecture, taste, or acceptance bars; the cost is that my throughput caps at my own judgment bandwidth.
- I price the worst case first — losses outweigh gains in my ledger; the cost is missing some high-upside bets.
The site's publication, authorship, and revision rules are documented separately in Method. They are not presented as personal experience.