AI runs it. It does not judge it.

An AI agent drives the pipeline. Fixed scripts build the board, independent checks and simulations judge it, an engineer signs it off. A board that fails a check cannot leave.

Build, check, fix, repeat.

Follow one request through the loop. The agent only acts through scripts; the gates only read the board file.

Example runRequirements in, release out, with one failed gate on the way.

One real board, straight from the pipeline.

Test run 10, a SpaceWire to SPI bridge. Every image is an output file of the build, not a mock-up.

Part of the generated schematic: the SN65LVDS049 SpaceWire transceiver, its connector and the 100 Ω terminations
Schematicgen_sch.py · ERC 0 / 0
Top copper: the FPGA fan-out and the SDRAM bus as 45° bundles
Copperroute.py · DRC 0 / 0 / 0
Thermal map of the routed board
Thermal model0.73 W · max 34 °C
3D render of the released board
ReleaseLVDS 100.0 Ω · SE 50.1 Ω

How it was routed →Open it in the viewer →Release files · zip, 2.9 MB

Eight stages. Each ends in a check.

The next stage starts only when the check passes. ● an engineer approves the stage.

StageThe agentThe scriptsMust pass
Inputs ●Turns your requirements into an input sheet; proposes parts with alternatives.Impedance and stackup tables for RF and fast boards.Engineer approves the inputs.
Feasibility ●Builds test patches of the densest areas.KiCad DRC on each patch; space and power budget.Patches pass DRC.
Schematic ●Writes the schematic generator.gen_sch.py writes the KiCad schematic.ERC 0 / 0; pins match datasheets.
Board setupWrites net classes, rules, stackup.proj_setup.py writes the project rules.Every custom rule fires.
Placement ●Places parts by script.place.py, from one geometry file.Courtyards clear; renders reviewed.
Routing ●Scripts buses, pairs and power.route.py writes copper; router output frozen.DRC 0 / 0 / 0; layout and RF lint 0 FAIL.
VerificationRuns the remaining gates, reads every report.Review, footprints, sourcing, thermal, RF, SPICE.All nine gates PASS.
Release ●Exports the package, writes the report.Gerber, BOM, placement, PDFs, version tag.Self-test rebuilds the board byte for byte.

What the AI can and cannot do.

Can

  • Write and change build scripts
  • Run tools, linters and simulations
  • Read reports and propose fixes
  • Propose parts, with alternatives

Cannot

  • Hand-edit the board: all copper comes from scripts
  • Waive a failed gate
  • Relax a rule for one board: rules are shared, and the self-test rebuilds every board when one changes
  • Release without an engineer's approval

Why it holds.

Reproducible

Fixed seeds and frozen router output. Anyone can rebuild the board and rerun every check.

Replaceable model

Rules, tools and gates live in the repository, not in the model. Change the model; the checks stay.

Learns by rule

Every review finding becomes a linter check with a test, applied to every later board.

A check proves the design meets its rules and models. Bench measurement proves the board: the first boards are ordered through the pilots. Every gate in detail →

Quote

Send your requirements.

A fixed price and date within one working day.