PCB CAD, Rebuilt for Touch

Professional PCB design is one of the last serious engineering disciplines with no credible touch-native tool. The iPad is a precise, pressure-sensitive, direct-manipulation instrument that professionals happily use for illustration, 3D sculpting, and music production. But for board layout, the assumption has always been: real CAD means a desk, a mouse, and thirty years of accumulated desktop interface.

We wanted to know whether that assumption is a law of physics or just a habit.

Why isn’t this already easy? Because PCB CAD is deceptively deep. Under the drawing surface sits a stack of hard problems: exact geometry with manufacturing tolerances, design-rule checking, footprint and symbol libraries covering hundreds of thousands of real parts, netlist semantics, multi-layer routing, and export formats fabs will actually accept.

Incumbent tools carry decades of that accumulated knowledge. A touch-first challenger doesn’t get to skip any of it. A beautiful sketching app that can’t produce fabricatable Gerbers is a toy. Historically, rebuilding that stack was a multi-year job for a funded team, which is exactly why nobody did it for a tablet.

Start from interaction, not from porting. Touch and Pencil are treated as the primary instruments: components are placed and nudged directly, traces are drawn as strokes that snap to deterministic, rule-clean geometry, and the interface is built around what fingers and a pencil are actually good at rather than around menu trees.

Everything under the surface stays rigorously boring: exact-arithmetic geometry, a real rule engine, real netlists. The novelty budget is spent entirely on the interaction layer.

The equation changed on three fronts at once:

• Open EDA data. Open-source footprint and symbol libraries, vendor CAD models, and openly documented formats mean the parts-library problem, once a moat measured in staff-years, can be mined, converted, validated, and continuously ingested by automated pipelines instead of hand-built.
• Datasheets as geometry, not pictures. A zero-dependency PDF compiler, its own DEFLATE decoder included, reads land patterns straight out of a datasheet’s vector drawing commands. Deterministic, no OCR and no model in the loop, on work that used to be pure manual labor at library scale.
• Modern client hardware. Current tablets have desktop-class silicon. The compute for real-time rule checking and interactive routing is already in the user’s hands.

None of these alone makes touch-native CAD viable. Together they collapse the team size the problem used to require.

A working touch-native layout environment, backed by an automated ingestion system that continuously harvests, converts, and quality-checks component libraries from open sources (footprints, symbols, land patterns, datasheet-derived pin data), plus a placement engine that validates itself by attempting the route: dropping an MCU offers its support circuitry from 3,337 bundled per-chip templates, and a bounded maze search runs as a preflight to prove there is room before the placement commits.

The pipeline runs as a set of coordinating agents that hand work between themselves through a written contracts registry, with a single-owner rule so two never touch the same artifact. Every push is gated by eight regression checks, each baselined so it fails only on a new defect and ratchets downward as things improve. It does work that used to be a librarian department.

Plenty, and they were informative:

• Naive touch translation fails immediately. A finger is roughly forty pixels wide; a trace is three. Every interaction had to be redesigned around snapping, gestural intent, and magnification rather than pointer precision.
• Automated library conversion is a swamp of edge cases. Real-world source libraries are inconsistent, subtly wrong, and occasionally lie about their own units. A large fraction of the pipeline is validation and quarantine, not conversion.
• AI-suggested routing without a deterministic geometry layer underneath produces plausible-looking boards that violate manufacturing rules. The lesson stuck: AI proposes, exact geometry disposes.

Active and unfinished, but past the existence-proof stage: boards can be laid out with a pencil, against real parts, with real rules, on a device that fits in a bag. The ingestion fleet keeps growing the component corpus without human attention. The remaining work is depth: routing quality, rule coverage, fab-export breadth. Not feasibility.

If one small lab can stand up a credible attack on professional EDA, a category owned by companies with thousand-person engineering organizations, then the moat around every legacy professional tool category is thinner than its owners believe. The moats were mostly accumulated manual labor: libraries, format handling, edge cases. That labor is exactly what AI and automation now compress.

Notebook The interesting boundary isn’t “can a tablet run CAD.” It’s “how much of a legacy software empire was actually just headcount.”