// writing & talks
I write and record about the parts of building AI systems that don't make it into the demo — retries, rate limits, and what "it works on my machine" turns into at production traffic. I also make it for whoever's paying attention on The Black Tech Experience.
// latest post
Retries aren't free — a look at how retry amplification wastes GPU capacity during inference (including a real 48-retries-per-failed-request stat from Microsoft's own May 29 Azure OpenAI incident), and why pacing demand instead of rejecting it can turn excess load into waiting time instead of wasted compute.
GitHub's August 17 outage is a glimpse of what happens when demand becomes reactive faster than supply can scale. Why retries, jitter, and circuit breakers aren't the whole control loop once agent traffic enters the picture — and what pacing at Layer 7 actually looks like.
Also wrote this for EZThrottle, a multi-region HTTP proxy that races outbound API calls across regions — the way Cloudflare routes inbound traffic, but for the requests your app makes to things like OpenAI and Stripe.
Why client-side retries fall apart during a partial regional outage, and how region-racing outbound requests across providers — the way Cloudflare routes inbound traffic — turns a 3am page into a non-event. Includes the real math on what a DIY retry queue costs once it needs its own team.
// on video
Part of an ongoing series breaking down no-code/low-code automation builds in under five minutes — for The Black Tech Experience.
Wrote, shot, and edited a direct-response ad for DocFetch aimed at accountants and bookkeepers chasing client paperwork.