Letting a router pick the cheapest healthy provider for an open-weights model is a good uptime decision and a silent jurisdiction decision. What I found when a customer asked which country processes their data.
Writing
Writing
Notes on systems, infrastructure, and the plumbing under AI products.
4 of 4 posts
- OpenRouter load balancing vs. data residency: three config keys I had never set
- Two of my four blocklist checks had been silently failing for days
Instrumenting a system is most of the work. Surfacing what it records is the remaining 20%, and that is where almost all the value lives. A short postmortem from running 200K+ sends a day.
- Gamified ASD Intervention: Aiding Emotional Development
A comprehensive look at a gamified approach utilizing HCI technologies to aid emotional development in children with ASD, featuring a computer vision pipeline with OpenCV and ResNet.
- Why perf, tshark, and valgrind Miss the Point: Measuring Latency in Multi-Threaded Applications
Traditional profiling tools measure CPU time spent, not response time. In multi-threaded applications, this distinction is critical, and often misunderstood.