Writing
Ideas, guides, and lessons from building enterprise software and integrating AI in production.
AI Engineering
Developer Experience
System Architecture
AI worms at home: slow does not mean harmless
October 14, 2026 · 10 min read15 TV boxes, a 26B model, and an uncomfortable question: how close are household AI worms? What Anthropic demonstrated, and what it projected.
Harness learning: is the model to blame?
October 12, 2026 · 10 min readHarness learning aims to improve AI agents by revising their workflow. What the paper shows, and what to check before switching models.
The bot isn't enough: why we built an MCP server
October 7, 2026 · 10 min readWhy we built an MCP server for MaltaClean's operators and admins, and what permissions, OAuth and actions with real consequences taught us.
Jev in Production: 703 Turns Later, Still a Complement
October 2, 2026 · 12 min readAcross 703 real WhatsApp turns, Jev responded in 128 of 131 main-classifier failures: new shadow-mode signals, not a proven replacement.
The endpoint that took 29 seconds (and nobody complained)
September 26, 2026 · 7 min readAutopsy of a ~29s production request: the AI call, the PDF render, and ~20 R2 downloads hidden inside an email — and the pattern that got it under 500 ms.
try/catch won't help when the process is already dead
September 26, 2026 · 6 min readA production FatalError at 30 s: the AI SDK was waiting for 60, PHP killed the worker first. The timeout hierarchy every backend needs.
We tested cheaper models and decided not to migrate.
September 24, 2026 · 6 min readA migration promised to simplify MaltaClean's concierge and lower the cost of each classification. Eight test cases showed why the savings weren't enough.
The model interprets; code authorizes actions.
September 23, 2026 · 6 min readA cleaning enquiry was mistaken for a job application. MaltaClean taught me why AI agents can interpret intent, but code must authorize actions.
The AI Model That Can't Write (and Why I Want It in Production)
September 19, 2026 · 9 min readI tested Jev, TypeSafe's first System One model: no text generation, just typed decisions. 292 cases, blind review, and 94.2% accuracy classifying a real cleaning company's WhatsApp.