I build systems that run while you sleep.

Senior Solutions Engineer — AI Automation

Enterprise AI company · remote

• Designed, security-reviewed and shipped a 24/7 autonomous LLM agent that owns a judgement-heavy operational workflow end-to-end — 0 errors across 192 production runs, ~75% less manual effort, ~€1.71 per unit of work
• Architecture: buy the runtime, build the thin brain — a managed agent platform for the reasoning loop and connectors, a thin layer on top for the rules, the learning store, the trigger and the metrics
• Earned autonomy the honest way: shadow mode → active mode, plus a hard escalation gate — security, PII, safety and ambiguity go to a human untouched, and the agent never creates, closes or overrides human-owned work
• Moved it from polling to event-driven (webhook → API Gateway → Lambda) — near-instant and cheaper; idle runs eliminated, volume self-tuned ~75/day → ~8–11/day
• Hardened the path: HMAC-SHA256 signature verification, secrets in a managed vault, no public function URLs, and loop-safety so the agent’s own writes never re-trigger it
• Wrote the third-party connector security assessment that gated its own production launch — found a self-asserted certification, undisclosed OAuth scope and an out-of-region LLM sub-processor; issued a sandbox-first verdict, not an opinion
• Shipped LLM spend governance that auto-disables a key at a hard cost or time cap, running at $0 and zero model tokens (admin APIs only)
• Earlier in tenure: serverless AWS ⇄ service-desk integration (manual escalation down 60%), analytics platform migration (€30–60k saved annually), €4–7k/year in licenses recovered