I ship the application, the layer underneath it, and the experiment that proves it worked.
An MCP server over a production commerce BFF
The storefront ran Next.js over a Node backend-for-frontend — the layer where the company’s own engineering actually happened. I wrapped nearly its entire API surface as MCP tools so coding agents could select what they needed and exercise the service directly in testing. Then I drove adoption across the team.
3 senior engineers adopted
in-house, before the frameworks
Server-side experimentation, end to end
Two hidden features — a guided filter and a video pill — shipped as server-side A/B tests: variant assignment and branching in the application layer. Real state, real cache interaction, no flicker. The interesting engineering question is “how would we know” — not “can we build it.”
seven-figure projected impact
0 client snippets
himal — a procedural Himalaya in Rust
A quiet builder on real satellite terrain: every valley an 8 km window of the actual Himalaya, villages grown from 12 documented settlement grammars. Implemented by Claude against a spec I wrote and froze — agent-driven development with the bar held by hand.
442 valleys · 8 km each
9,566 named places
A local music-video pipeline + its quality gates
AI song in, finished music video out: FLUX.2 klein keyframes, LTX-Video animation, Kokoro voice, ffmpeg assembly — fully local on one 16 GB card, serial by necessity because the models can’t share it. Cut boundaries come from the measured beat grid; every in-point ends before likeness drifts.
6,368 frames, plan-exact
gate: 0/6 → 6/6 caught
owned the PDP’s entire image & video gallery · shipped across PLP, PDP, and the rental build — product customization, rental bundles · core web vitals work on LCP and CLS
The four records above are the survivors, not the collection. No single folder is the point — the pattern is: whatever appears at the frontier gets pointed at a real problem the same week. Twelve that earned their disk space:
counted, not estimated — every line traces to an artifact on disk: a render, a benchmark row, a merged PR, a photograph.
Before software: six years in the US Army as a CBRN specialist — last tour Atsugi, Japan. Served part-time straight through a computer science degree at the University of Houston. A scheduling problem I don’t recommend, but don’t regret.
Working style: make the system legible to the agent, then hold the bar yourself. The commerce work and the AI work aren’t two interests — they met in the same codebase, and they’ve stayed together since.