Production AI reliability, built on a decade of Rails and Ruby

I make production systems reliable; these days that mostly means AI - pipelines, agents, and the evals that keep them honest. The current proof is a pipeline that reads 500+ news articles every morning, writes and fact-checks its own briefing, and sends it with no human in the loop - and most of the work isn't the model, it's the schema validation, the retries, and the observability that keep it from quietly breaking at 6am. The goal each time: a system that says "I don't know" instead of guessing.

The reliability work is older than the AI. At Hivebrite I scaled a Rails platform serving over a million users, cut €160k a year in infrastructure, and took exports that used to die overnight down to minutes. At OmbuLabs I unblocked a CI pipeline that had been red for years and deleted 715k lines of code. The fixes land wherever the failure is: a performance fix merged into Rails, a lock-free fix merged into NextDNS's Go resolver, fiber-safe hub storage merged into Sentry's Ruby SDK, production-safe streaming proposed to Sinatra. AI systems break in newer ways, not different ones, and that's where I point the discipline now.

I want to know why we're building something, not just what's on the ticket, and I'd rather test a theory against real data than blindly apply it. I write about what breaks and what I learn on my blog.

Before any of this: a childhood in French lycées across four continents; I haven't stopped moving since. More on background.

I take this on as fixed-scope engagements, priced up front: a €2,500 diagnostic, and you know where it fails and what each failure costs, so the call on what to fix first takes a sentence; a €6-8k one-week sprint, and the worst of it is gone, with a before/after number you can put in front of anyone; an audit from €18k, and you keep guardrails that catch trouble before your users do, with the numbers to show it. Got a system like that? Drop it in my inbox. More in the CV, or ask the assistant here about my work (or point your own LLM at /llms.txt).