We build instruments — not platforms, not visions, not promises. Each one solves a specific problem for a specific kind of engineer. We ship them, charge for them, and improve them.
The past three years of AI tooling were built for a specific kind of person — a knowledge worker writing prose, code, or queries that live entirely inside a screen.
But the engineers who actually move the physical world — embedded developers, roboticists, IoT integrators, industrial control programmers — were never the audience. Their code runs on microcontrollers with 256KB of RAM. Their bugs surface only at 3 AM in a warehouse. Their tools were built before transformers existed.
We don't try to be a platform. We don't try to be a model. We make precise, opinionated instruments — one at a time — for engineers whose problems the rest of the AI industry overlooks.
Every PhyCyber tool is shaped by the same principle: do one thing for one kind of engineer, and do it better than any general-purpose AI can.
A local static analyzer for embedded C, FreeRTOS, and ISR-context mistakes.
Browser-based diagnostics for sensor data streams. Drift, noise, alignment — solved by an agent that understands signals.
A protocol for AI agents to query and reason about physical devices — not just talk about them.
I spent years moving between worlds most people pick one of. AI agents at Alibaba — writing code that fixed code. MEMS sensors and nanopore sequencing at Tsinghua — studying how physical signal becomes digital truth. Smart buildings, IoT firmware, warehouse pick-to-light — actual systems in actual factories, seven years ago, when none of this was fashionable.
PhyCyber is what that career converges on. The space between clean software abstractions and messy physical reality is where the next decade of AI tools needs to live. I'm building one tool at a time, for the engineers who already live there.
Occasional essays on physical AI, engineering tools, and the strange overlap between sensor physics and software design. No build-in-public diaries. No daily updates. Just the things worth saying.
The dominant AI coding assistants were built for a kind of code that doesn't have to survive contact with the physical world. The gap is wider than it looks.
Read essay →Foundation models keep getting smarter. The signals they reason over keep being garbage. We need to talk about this honestly.
In progress →