About
I turn probabilistic systems into products people trust.
I'm 王乙锟 (William Wang) — an AI Product Manager who builds the whole stack. I work where models meet people: turning a probability distribution into something a person can understand, rely on, and pay for.
In practice that means workflow runtimes where code is the source of truth, LLM backends that hold under load, and retrievers built for agents instead of eyes. I write the Rust, the Go, and the TypeScript — and I care about the type, the latency, and the empty state as much as the model.
I lead with intuition and build fast — and I break new ground more often than I follow a map. This page is the honest version of how I think and what I've built.
Prompts teach patterns, not facts
Hard-coded facts rot. Every version of my workflow engine moved a fact out of the prompt and into a runtime signal — function calls, context, a catalog schema — so the prompt only teaches the shape, and the truth arrives live.
Code is the single source of truth
The visual graph is a read-only projection of the code, never a second authority. One arrow, not two. Bidirectional sync is a merge conflict you've agreed to have forever.
Reliability comes from structure, not smarts
The bet behind my current engine: keep the model on rails — tight constraints, small pieces, type checks — and a small model can reliably do work that today looks like it needs a frontier one. Don't buy reliability with a bigger model. Build it into the structure.
Keep the core framework-agnostic
Business logic outlives whatever framework wraps it this year — especially when the frameworks get concurrency wrong. The part that survives load should not depend on the thing most likely to be replaced.
Fail loud, design for the tail
Silent fallback is how a probability distribution lands on a user as a surprise. Design for the tail first, name the cost you paid, and make the wrong answer cheap to correct.
Intuition-driven
Intuition is the instrument. From a single set of errors I feel the systemic problem before any metric would show it, and I steer with small, agile tests rather than big suites. I'm fast and I break new ground — and honest that thorough testing is my weak hand, so the colleague I want most is someone great at finding the problems I'd miss.
Whatever you name, I can build it
TTS, STT, AI painting, large language models, face-swap, retrievers, workflow engines — point at any layer of the stack and I've shipped something there. Range is the ability I trust most; the through-line is making a model into something a person can rely on.
People-first (以人为本)
INFJ. I build where a model meets a person, and I treat that person — and the team shipping to them — as the point, not an afterthought. A system that's correct but illegible has failed.
Forge — plain language → runnable LLM functions
Compiling plain language into typed, runnable workflows that SMALL models can get right — by keeping the model on rails (tight constraints, small pieces, type checks) instead of leaning on raw smarts. The grand version of everything before it.
High-concurrency LLM backend
A four-stage pipeline (grounding → intent → fan-out routing → polish) with a framework-agnostic core and an offline regression harness. Tuned for the tail, not the average.
Retrieval · monitoring · voice
A retriever built for agents (recall + cross-encoder rerank, scored with paired tests); a real-time intelligence dashboard; voice-first coding on AR glasses. Different surfaces, one obsession: make the model trustworthy.
Generative-art product · internship
A solo exploration shipped during an internship — generative anime-art tooling with a Discord-native product loop.
Meaning over money
I choose work for whether it means something, not what it pays. AI isn't a career to me — it's how I live.
Doers, not red tape
No bureaucracy. I want people who are here to make the thing — judged by what ships, not by process.
A team that feels right
Atmosphere matters to me as much as the work. The best teams I've been on were people I'd choose to spend the day with, building something we believed in.