Text to PCB

From prompt to board, in four steps

1.You describe the board

Write what it needs to do in plain English — “an ESP32 sensor board with USB-C power, a LiPo charger and two Qwiic connectors.” No schematic to draw first.

2.AI drafts parts and a netlist

A language model does the judgement half: it chooses a suitable microcontroller, regulator and connectors and proposes how they wire together. This is what LLMs are genuinely good at.

3.A deterministic engine builds the board

Everything measurable is handed to a deterministic engine: it places the parts on a board with real dimensions and routes the copper across layers. Routing is a geometric search against hard constraints, not text generation.

4.It checks the rules and exports

A design-rule check runs, the unrouted nets are reported by name, and the result exports to Gerber, drill, pick-and-place and BOM — the files a fab actually accepts.

Why text to PCB actually works

The reason generated boards are usable is that the AI is not asked to do the part it cannot. A language model has no way to measure whether a trace clears a pad by 0.15 mm — so that job goes to a deterministic engine instead. The AI decides intent; algorithms verify the physics. That division is what turns a prompt into a board a fab will accept, rather than a plausible-looking picture.

The questions worth asking

Is the board just made up by an AI?

No. The AI proposes intent — parts and connections. The placement, routing and design-rule check are done by a deterministic engine where every result is a calculation you could verify by hand. The AI decides; the engine proves.

Will it be the same every time?

The routing is deterministic: the same design produces the same board, and it does not drift when an underlying model is updated. That reproducibility is the point of keeping the two halves separate.

What if it cannot finish a board?

It tells you. Instead of a completion percentage that hides failures, the check reports the exact nets it could not route, by name, so you know precisely what to adjust — usually a placement change.