Notes from production
Long-form breakdowns of systems I've actually shipped. No listicles, no hot takes.
Unused permissions, priced in dollars
A license-saving feature end to end: live-computed findings, one action that becomes removal tasks, and action records engineered to survive the nightly data refresh.
Route to the machine that already knows the conversation
Most AI serving cost is re-reading conversations a computer has already seen. Nostos sends each request where its past already lives. Plus a public arena for racing routing strategies.
A language model from scratch, three times
The same transformer at three levels: PyTorch with no shortcuts, dependency-free C++ with tests green, raw GPU code in progress. What changes each time you rebuild it.
SoD for AI agents: toxic combinations in autonomous systems
Five risk types across agents, owners, chains and credential groups. AGENT_ONLY to INVOKER_COMPOSITE, in a deterministic three-pass engine.
From 21 hours to 3 minutes, via 2 seconds of compute
A compliance check re-examined the company hierarchy once per rule. Resolving it a single time collapsed a day of computing into minutes. 2 seconds of compute, the rest database writes, identical results.
Exports without an upper bound, in 7 MB of memory
The 5-stage bounded pipeline behind unbounded Excel exports. 15M+ rows proven in production. POI SXSSF, S3 multipart, streaming ZIP.