Writing
Write-ups from production.
Case studies, a re-runnable benchmark, and field notes from real systems under real load. Every claim in them traces to something you can check: public data, a public repo, or a decision written down when it was made.
I audited 119 repos for AI-agent readiness.
I run a public leaderboard that scores open-source repos for AI-agent readiness. As of July 2026 it holds 119 repos, and 41 percent have nothing that tells an agent how to work. Here is the full distribution and what to add.
Read itThe real-time bug that only appears once you run two pods.
A game studio's live dashboards froze for half their players the day we added a second server, and the logs showed nothing wrong. The network took the blame first and was innocent. Here is where the data actually got stuck, and the fix that unstuck it.
Read itThe CI gate that actually holds.
Making a deploy wait for the checks to pass sounds like a one-line job. It is not. The two obvious ways to wire it up both fail without a sound. Here is the one that holds: put every check in a single list of steps, with the ship step last.
Read itThe control plane managed cloud hides.
Managed cloud is a control plane with a nice button on it, built by someone, running underneath. A game studio hired me to build their whole hosting product on plain rented servers, with none of that handed to me. So I built the control plane by hand: a sequence that stands up a server, joins it to the pool, and rebuilds it from scratch when its first boot fails, taking a one-in-three failure rate down to a few percent.
Read itWhen does self-hosting actually pay?
Break-even numbers from a real production benchmark, with the token logs behind them: measured per-call API costs from roughly a dollar to $4.92 per thousand calls, GPU options from $160 a month rented to $60 a month owned, and the honest answer about when self-hosting actually pays. Formula included, so you can run it on your own volume.
Read itWhat actually breaks in vibe-coded apps.
49 verified community threads about Lovable, Bolt, v0, Replit Agent, and Cursor apps failing in production, classified: the top failures are cost lock-in, then security, then agents breaking their own code, with performance crashes near the bottom. With the 8-point pre-launch check and the full dataset.
Read itWhere a demo ends and production begins.
A demo proves an idea works once. Production is where it has to survive real cost, real failure, and real load. The three places software breaks when real users arrive, and how to engineer against each, with the measured proof behind every claim.
Read itA self-hosted model matched cloud quality at a tenth of the latency.
A reproducible benchmark. 1,200 vision-analysis calls across six backend arms, three cloud models, one also run with extended thinking off as a control, and two self-hosted, on the same scenes with two independent LLM judges. A self-hosted 30B model matched the best cloud model's quality within statistical noise while answering in 0.77s, five to ten times faster, at no per-call cost. Raw data and the runner are in the repo.
Read itA one-page concept, taken to a multi-region game-hosting platform.
A case study. Over a six-month engagement I led the design and build of MetricHost, a multi-region game-server hosting platform, from a one-page concept to production. k3s, Cilium eBPF, a Go control plane, idle-server hibernation, and the honest cost call that ended it.
Read itWhen a vendor moves the floor overnight
When a single hosting vendor repriced its machines overnight, it broke the assumption a whole business was built on. A write-up on doing honest unit economics: verifying per-GB-RAM cost at the source, building a portable multi-vendor mix, separating RAM efficiency from instance price, and telling a client the real number.
Read it
If one of these sounds like the problem on your desk, describe it to me. You get a straight answer about where I'd start, and it costs you nothing to ask.
Tell me what's breaking