Things you can check yourself.

Production migration · my engineering work at Mediavine

Moving a 100k req/s auction server to Rust without stopping it.

S2S is the real-time bidding service behind Mediavine's ad auctions. Every page view gives it a few milliseconds to run an auction before the opportunity is gone. I led its migration from Node.js to Rust as an incremental replacement rather than a rewrite: production traffic ramped from 2% to 100% in measured stages, and each stage had to earn the next one with real production data.

The correctness gate was deterministic. Generate production artifacts, ingest them in development, make the bytes match. If the bytes didn't match, it didn't ship.

Feature work, code reviews, and production support never paused during the 43 days it took to reach full traffic.

The efficiency gains turned into money fairly directly: roughly $87.5k in additional publisher revenue in the first 24 days after launch, on top of thousands of dollars a day in infrastructure savings.

Shipped products · NG Tech

Apps that made it through App Store review, repeatedly.

Two apps are live: Unearth, a quiet offline excavation game with no backend at all, and Let's Dance Robot, which has been through fifteen releases since December 2025. Getting one app approved is luck. Getting the same app approved fifteen times in a row is a process, and the process is the part worth hiring.

The tooling behind it handles build triage, device verification, screenshots, and submission artifacts the same way every release, on both stores now. It drafts and it checks. It does not decide — the process is written up here.

• Versioned privacy surfaces — every shipped version pins its own privacy policy URL on this site, so the disclosures a build was actually reviewed against stay live permanently.
• Remote operational controls — the ad stack reads a kill switch served from this same Worker, so production levers live in one place I control.
• Two apps, under twelve hours — the most recent pass took two feature ideas in two different apps from decision to live on the App Store in under twelve hours, both through review.
• Human approval at the end — release calls, reviewer replies, and privacy disclosures are mine every time. Those are the expensive things to get wrong.

Live platform · NG Tech

The publishing pipeline serving this page.

The Insights hub is an agent-driven research pipeline. Agents collect and synthesize primary sources into Notion, an editorial gate decides what publishes, and a deterministic publisher turns approved entries into committed Markdown and a typed manifest baked into a Cloudflare Worker at deploy time. No frontend framework, no client-side JavaScript, no runtime data fetching, and effectively no hosting bill.

It is also the clearest demonstration of how I think AI should be wired into a real process: the machine does the reading and the drafting, a person decides what goes out, and the output is labeled so nobody has to guess.

• Server-rendered TypeScript — every page is a typed template string with inline styles, and no client-side JavaScript ships at all.
• OAuth-protected MCP ingestion — hosted agents write signals through a separate Worker exposing Model Context Protocol tools, with dedup enforced server-side rather than trusted to the agent.
• Deterministic publishing — reruns with no upstream changes produce no diffs, so the whole pipeline is safe to re-run from CI on a schedule.