Ten years of making AI systems that survive production.

I started where failure is expensive: defence-grade video surveillance analytics at BEL and eInfochips — multi-stage vision pipelines, embedded constraints, systems that had to run unattended.

Alongside, freelance ML and data projects taught me the other half of the job: scoping, shipping, and owning outcomes without a team to hide behind. Production reliability stopped being a virtue and became a habit.

Systems that could explain themselves came next — under a regulator's clock.

The LIBOR transition put NLP inside a regulatory deadline: thousands of financial contracts, clause-level extraction, and zero tolerance for unexplainable output.

I learned to design for governance — confidence-gated automation, human review queues, audit lineage from source clause to decision. The lesson stuck: in the enterprise, trust is an architecture concern.

I took that bar into a domain where the edge cases walk in the front door.

Six years deep in one domain — European hospitality. RAG support platforms, multilingual email orchestration, recommendation engines, review moderation, computer-vision service intelligence — different modalities, one industry, end to end.

Domain fluency changed how I architect. The hard problems were never just models — they were workflows, data ownership, and what the people on the floor actually do at peak hour.

Ten years of patterns deserved a platform of their own.