I spent a decade on the pastry side of Michelin kitchens — eight years at a two-star in Taipei, two at a three-star in Tokyo. The same discipline that builds a flawless dessert service now goes into production LLM systems with verifiable guarantees. I'm an architect first: the decisions are mine, the implementation velocity is the AI's.
In a Michelin kitchen, the pass is where every dish is checked before it leaves — nothing goes out wrong. I run engineering the same way. This is the through-line, in order.
No culinary school — I walked into the kitchen on passion alone, starting front-of-house as a server and working my way up to pastry section manager. The hardest part of scheduling was carving out time, in the middle of the daily rush, to face a blank roster against every cook's time-off request. So I let a system finish the hardest 80% first, and hand the last — and most critical — 20% of fine-tuning back to the manager.
Output measured against the highest bar there is. Japanese business fluency, and a culture of precision under pressure that maps cleanly onto production engineering.
I built the first version of the scheduler learning as I went, then proudly took it to a senior-engineer friend for review. He glanced at it and said, 'Your backend is a mess — it runs, but you could never apply for a job with this.' That was the moment I understood a real project has to hold backend discipline — and why architectural judgment now stays firmly in my hands.
Training robots to collaborate with humans in professional culinary environments. The one place where ten years at the station and a new career in AI point at exactly the same thing.
Two production systems built on the same conviction: an LLM may interpret messy human input, but it never decides the outcome — a deterministic core keeps the guarantee. Held together by a culinary narrative only I can tell.
A LINE messaging assistant running for my family's homestay business in Yilan. An LLM acts strictly as an input-boundary parser — emitting structured JSON (intent + slots + confidence) — while a deterministic quote engine keeps every pricing answer truthful. Real guests, real money, real failure modes.
A staff-scheduling service whose deterministic engine deliberately stops at an explainable 80% draft, surfaces what's still wrong through warnings, and hands the last 20% to the manager through a bounded natural-language refine loop. Built on the exact problem I lived for a decade.
Honest about where I stand: two systems shipped, and a clear path from here.
I want to train humanoid robots to work alongside people in professional kitchens — not to automate the chef away, but to stand beside them like a strong commis and share the load. What pulls me toward this is what I saw in the kitchens I worked in: so many people walk into this trade on passion and a dream, only to be worn down, bit by bit, by punishing hours and a life knocked out of balance — and the industry keeps getting harder to staff. I believe a lot of that heavy, grinding work could be handed to robots — so the people who still love this craft don't have to burn themselves out to stay in it. The years in the kitchen and the engineering I've built since — very few people in the world can stand at both ends of that scale. This is the career goal I'm setting for myself, and the one thing I'm most certain I was meant to do.