Projects
Development decisions, tests, and lessons from the PCB Automation project.
12posts
ProjectsPCB AutomationReading PADS ASCII coordinates as 1nm made the parts, the traces and the board all 1.5× too large together, so every internal consistency check stayed green while the object was wrong.
ProjectsPCB AutomationA backslash that invalidated an entire file, pad fields shifted by one slot, and a 3D model rotation we guessed wrong three times - what we hit while moving 1,214 footprints and 857 symbols into a KiCad library.
ProjectsPCB AutomationThe 45° rule, staircases and antennas - none of which DRC catches - each needed a checker of its own, and the per-layer routing directions had to be injected straight into the DSN file.
ProjectsPCB AutomationIf the measurement changes the subject, it is not data. How we caught a check command rewriting the board and a fixer script shaving trace widths to turn DRC green, then measured our own shipping boards for the real numbers.
ProjectsPCB Automation12.9 bends per net and a detour ratio up to 3.0× said what the gates would not: a 16-channel board was being crammed onto two layers. How we moved it to four and re-injected the per-layer directions.
ProjectsPCB AutomationThe longest stretch in automation is not the first 90 percent but the last single item. Ranking the metrics as shorts over unrouted over clearance, and solving a blocked spot with a transaction that moves the wall.
ProjectsPCB AutomationDRC 0, unrouted 0, shorts 0, measured five times with the same result five times. The closing round that took arbitrary angles to zero, plus the seven ways our own tools lied along the way.
ProjectsPCB AutomationThe metrics we measured passing is no evidence that the items we did not measure are correct. The schematic, label and angle gaps a review exposed after we called the board done, and the gates built to close them.
ProjectsPCB AutomationAn unrouted net is not a conclusion but the starting point of the next investigation. Measuring in coordinates, suspecting the tool, unifying the pass criteria, and finally putting the placement choice itself inside the loop.
ProjectsPCB AutomationMeasure the same board with two rulers and you get two answers, and the wrong one was always ours. Ten disagreements found by changing how we measured rather than by fixing the tool.
ProjectsPCB AutomationTen posts on PCB design automation, from generating the schematic to reaching DRC 0, collected on one page in the order the problems appeared.
ProjectsPCB AutomationEvery time we scaled an 8-channel board to 16 channels, we had to redo the schematic and artwork. To cut the repetitive work, we introduced AI automation and built a library and pipeline.
