PCB design review

AI can draw a board. That does not mean it will work.

GPT-6 Astra and similar agents now produce KiCad layouts that look complete at first glance. The boards we have seen so far would not be manufacturable, would not pass EMC, and often would not survive first power-on.

Populated control board with MCU, headers, and Ethernet

What we offer

A human design review for AI-generated schematics and layouts – so you get a board that can be fabricated, assembled, and operated. Not a campaign against the tools. The missing check before you spend money on prototypes.

Unpopulated PCB with QFN land pattern

Typical starting point

The board looks finished

An autorouter or agent connected every net. DRC is green. Decoupling, return paths, creepage, or a fab-ready stackup are still missing.

Footprints and the BOM do not match

Land patterns do not match the ordered part, pads are mirrored, or the agent used a generic package that does not exist.

Power and protection are incomplete

No real PDN, missing TVS or reverse-polarity protection, unclear current paths, traces too narrow for the actual load.

From the bench

Real boards, not a chat rendering

What we review

  • Schematic intent: function, net names, missing pull-ups, reset, boot straps, analog references
  • Footprints and land patterns against the datasheet, not the first library hit
  • Stackup, impedance, return current, analog/digital split, creepage and clearance
  • Power integrity: decoupling, ferrites, load steps, thermal bottlenecks
  • EMC-relevant routing: ground slots, antennas, USB/Ethernet, switch-mode supplies
  • DFM and assembly: paste, courtyards, polarity, test points, panelization
  • Mechanics: connector placement, height, enclosure, screws, keep-outs

What you get

  • Written findings by severity: blocks fabrication, field risk, or clean note
  • A clear statement of what would fail in reality – not just a DRC dump
  • A recommended fix path, ordered by effort and risk
  • Optional: we implement the corrections and deliver fab-ready data

What we need from you

  • Project source: KiCad, Altium, Eagle/Fusion, or Gerbers plus BOM
  • Schematic as PDF or native file
  • Target fab, volume, voltages, environment, and standards if known
  • What the agent was asked to build – not only what it produced

Common questions

Do you only review AI-generated boards?

No. The current influx comes from GPT-6 Astra, KiCad agents, and similar workflows. The same review applies to human-drawn layouts that need to go to fab.

Is a green DRC enough?

No. DRC checks rules someone defined. It does not check whether the net does the right job, whether the footprint matches the part, or whether the board can be assembled and powered.

Can you also fix the board afterwards?

Yes. The review ends with a prioritized correction path. If you want, we implement the changes and deliver fab-ready data.

Which tools do you accept?

KiCad, Altium, Eagle/Fusion, plus Gerbers, drill files, BOM, and schematic PDF. Native sources are better than fabrication data alone.

Send the board before you fabricate

Briefly describe what the agent started from and which files you have. We will tell you whether a review is enough or whether the layout should be restarted.

Or use the contact form with the review type already selected.