Synthwave PCB landscape
Self-Improving Local AI for Hardware

GerberGenius Pro

One sentence in. A manufacturable project out.

A one-of-a-kind local AI that takes plain English to PCB to 3D-printed enclosure. Type a description — get a KiCad PCB, 12 JLCPCB-ready gerbers, a parametric OpenSCAD enclosure, a print-ready STL, and Prusa gcode. Every artifact verified by the real tool. No cloud. No API. Running locally on a Mac.

View All Projects

What Is It

A fine-tuned Llama 3.1 8B model paired with 18 real industrial tools — kicad-cli 10, OpenSCAD, PrusaSlicer, FreeRouting — that generates complete PCB + enclosure projects from natural language.

How Does It Work

The AI outputs a JSON spec. Real compilers verify every artifact. If the output doesn’t pass, the exact error is fed back until it works. The model trains on its own verified successes — AlphaGo-style self-play for hardware design.

What You Get

87.5% pass rate on real kicad-cli DRC. Gerbers you can upload to JLCPCB today. Enclosures sliced to Prusa gcode. 11,994 real KiCad footprints. No API costs. Your designs never leave your machine.

One Prompt → Full Pipeline

From One Sentence to a Manufacturable Project

You type a description. The model outputs a JSON spec. A 12,000-footprint library expands it into a valid .kicad_pcb. kicad-cli runs the real DRC. OpenSCAD compiles the enclosure. PrusaSlicer slices the gcode. If any step fails, the exact error loops back until it passes.

  • KiCad PCB with real components — passes real kicad-cli DRC
  • 12 JLCPCB-ready gerber files you can upload today
  • Parametric OpenSCAD enclosure with standoffs matching the PCB
  • Print-ready STL verified by the OpenSCAD compiler
  • Prusa gcode sliced by real PrusaSlicer for the Prusa Core One L
gerbergenius-pro
> ESP32-C3 dev board with USB-C, 3.3V LDO, boot+reset buttons, LED, 45x30mm, 4 M2.5 holes
# resolving components against 12k footprint library...
# emitting .kicad_pcb + running kicad-cli 10 DRC...
PCB: 45.0 × 30.0 mm 2-layer FR-4
DRC: PASS (0 errors, 0 warnings)
Gerbers: 12 files JLCPCB-ready
# generating OpenSCAD enclosure...
Case: snap-fit lid + USB-C cutout
STL: compiled OpenSCAD 2021.01
GCode: 42 min PrusaSlicer 2.9.4
✓ project verified end-to-end

The Architecture

The AI is the architect. The tools are the draftsmen. The compilers are the inspectors.

☷

Real Tool Verification

  • kicad-cli 10 DRC on every PCB
  • OpenSCAD 2021.01 compiles every .scad
  • PrusaSlicer 2.9.4 slices every STL
  • FreeRouting 2.1.0 headless autorouting
↺

Self-Improvement Loop

  • AlphaGo-style self-play for hardware
  • Winners become the next training set
  • Promoted only if benchmark improves
  • Auto-rollback on regression — no humans
▭

Parametric Enclosures

  • Snap-fit or screw-mount lids
  • Standoffs matched to PCB mounting holes
  • Cutouts for USB-C, barrels, LEDs, glands
  • Vents, OLED windows, button caps, ribs
✓

JLCPCB-Ready Output

  • 12 gerber files per project
  • Drill, BOM, CPL, STEP 3D export
  • JLCPCB 2-layer rule checker built in
  • 11,994 real KiCad footprints, 132 libraries

Why It’s Different

Every other AI that touches PCB design either hallucinates or runs in someone else’s cloud. This does neither.

General AI
GerberGenius Pro
Generates code that looks right
Every artifact verified by the real tool
Output is a linter-pass at best
kicad-cli 10 DRC, real OpenSCAD, real PrusaSlicer
Invented part numbers
11,994 real KiCad footprints, fuzzy-searched
Static once trained
Trains itself — 79 versions, auto-promoted
Runs in someone else’s cloud
8B model on a Mac. You own the weights.
“Export your Gerbers”
12 JLCPCB-ready gerbers + STL + Prusa gcode

The Numbers

79 model versions. 226 projects generated and verified. One laptop.

87.5% DRC Pass Rate 24 benchmark projects
206k+ Examples real hardware training pairs
79 Model Versions v1 through v79
11,994 Footprints 132 KiCad libraries
145 Projects Shipped end-to-end verified
84 Real Gerbers JLCPCB-ready outputs
161 GCode Files PrusaSlicer verified
180 Real PCBs Parsed Olimex, HackRF, Adafruit

Output Ecosystem

Real files for real manufacturing pipelines

Design Files

☷

KiCad Schematics

.kicad_sch files

▣

PCB Layouts

.kicad_pcb files

▭

OpenSCAD

Parametric .scad code

◆

STL Export

Print-ready meshes

Manufacturing

⚙

Gerber Files

JLCPCB-ready outputs

☶

BOM Export

CSV with LCSC codes

☉

CPL Files

Pick & place positions

✓

DFM Reports

Pre-flight validation

🔒

Local & Private

Mac Studio M5 Max, 128GB RAM. No API. No cloud. No data leaves your machine.

⚙

MLX on Apple Silicon

Llama 3.1 8B, 4-bit quantized. LoRA rank 16, alpha 32, 24 layers. Fast inference without a GPU server.

💰

Zero API Costs

No tokens, no subscriptions, no per-query billing. 79 training runs and counting — all on-device.

🛡

You Own the Weights

The model on your machine is the same one that trained itself. No telemetry, no cloud sync, no lock-in.

Get Started

Three lines. That's it.

terminal
# install the runtime
$ pip install mlx-lm
# load the latest self-trained adapter
$ cd GerberGenius-Pro-v79
# one sentence in — a manufacturable project out
$ python run.py

The Stack

Llama 3.1 8B LoRA Fine-Tuning MLX M5 Max KiCad 10 OpenSCAD PrusaSlicer 2.9.4 FreeRouting JLCPCB Self-Play Real DRC Prusa Core One L

One Sentence. A Manufacturable Project.

A small local AI given real industrial tools and the ability to improve itself. No cloud APIs. No venture funding. Just a laptop, real compilers, and patience — and 145 verified end-to-end projects to show for it.

Explore All Projects →