a
AI placement & routing
Train models that understand full-board context — not just local routing heuristics.
§ 01
The premise
Modern AI can write code because the internet contains billions of examples of software. Electronics has no equivalent.
PCB designs exist across GitHub, GitLab, OSHWHub, EasyEDA, KiCad repos, Eagle archives, and manufacturer exports — but they are fragmented, duplicated, format-incompatible, and mostly unusable for ML.
CommonCircuits turns that mess into structured training data.
§ 02
What we are building
A normalized dataset of real-world PCB projects.
Input sources
Normalized outputs
KiCad projects
schematic graph
Eagle
.sch / .brd fileslayout geometry
EasyEDA / LCSC projects
component–pin–net graph
Altium-derived exports
footprint / pad / via geometry
Open hardware repos
routing topology
Fabrication artifacts
board constraints & manufacturability signals
fig.02 — machine-learning-ready representation
schematic → components, pins, nets, values, hierarchy
layout → board outline, layers, pads, vias, tracks, zones
constraints → stackup, clearances, net classes, keepouts
metadata → domain, license, quality score, dedup family§ 03
Why now
PCB design is one of the last major engineering workflows without a public foundation dataset.
AI EDA models need examples of how real engineers turn intent into manufacturable hardware:
input
schematic graph
+ constraints
+ constraints
model
placement
+ routing
+ routing
output
fabrication
package
package
Today, every team trying to build AI PCB tools has to start by scraping, cleaning, parsing, and deduplicating the same chaotic public data.
CommonCircuits makes the dataset layer reusable.
§ 04
Initial corpus targets
Mostly 2-layer boards — ideal for early layout models, design-rule reasoning, and schematic-to-board automation.
-
50K–150K
raw repositories
with PCB artifacts
-
20K–60K
schematic + layout
paired raw designs
-
5K–15K
unique paired
clean, after dedup
-
50K+
noisy pretraining
corpus of artifacts
-
500–2K
benchmark
high-quality designs
Harder domains — RF, high-speed digital, dense robotics boards, instrumentation, power electronics — are rarer in public data. That scarcity is part of the opportunity.
§ 05
Why it matters
CommonCircuits enables a new generation of AI for hardware.
b
EDA agents
Evaluate whether an agent can go from schematic to a valid, manufacturable board layout.
c
Design-rule learning
Decoupling, power paths, connector orientation, via usage, zone fills, trace topology — learned from real designs.
d
Synthetic augmentation
Generate controlled variants of real designs for reinforcement learning and simulation.
e
Public benchmarks
Open evals for schematic parsing, netlist matching, placement, routing, DRC repair, and manufacturability.
§ 06
Deduplication & quality
Open-source hardware data is extremely duplicated — forks, templates, tutorials, reference designs, versioned commits. The same board can appear hundreds of times.
// fingerprint
Electrical & geometric hashing
- ▸ netlist hash
- ▸ component multiset hash
- ▸ footprint multiset hash
- ▸ board outline hash
- ▸ placement graph hash
- ▸ routing topology hash
// quality flags
Every design is labeled
- ✓ parses cleanly
- ✓ schematic / layout netlist matches
- ✓ libraries resolved
- ✓ DRC status known
- ✓ fabrication files present
- ✓ license usable
- ✓ manufacturing metadata present
§ 07 · positioning
CommonCircuits is the CommonCrawl for circuits.
A public and proprietary dataset layer for training, evaluating, and deploying AI systems that understand electronics design.
§ 08
We're looking for
Collaborators across the open hardware, EDA, and AI ecosystems.
- data open hardware repositories & PCB datasets
- eng KiCad / Eagle / EasyEDA parser contributors
- research EDA & ML researchers
- industry hardware companies willing to share anonymized design artifacts
- capital investors interested in AI-native engineering infrastructure
Contact the team
sava@dayworkx.com
/ one-line pitch
CommonCircuits turns the world's fragmented open PCB designs into the training corpus for AI-native electronics design.