Physics-drivenAI for electronics/ design.
Trained on the world's largest electronics corpus — not the world's largest chat corpus.
Not a chatbot.
An inference engine for hardware.
GerberGPT is a transformer pretrained on real electronics — parts, packages, schematics, and pad stacks — then fine-tuned to reason about circuits the way an engineer does. It doesn't retrieve text; it designs.
Schematic Inference
Reads a partial schematic and infers the missing nets, bypass strategy, and connectivity — completing intent in seconds instead of hours.
Component Replacement
Swaps any part for a pin- and parameter-compatible alternate, reasoning over fine-tuned parameters to keep the circuit electrically valid.
PCB Feature Design
Generates board features — planes, decoupling, length-matched buses — from intent, opening design spaces human iteration rarely reaches.
Footprint Generation
Produces accurate landing patterns and courtyards straight from datasheet geometry, vectorized so placement and DRC stay consistent.
Layout Tracing & Routing
Infers trace topology from the netlist while respecting impedance, clearance, and manufacturability — routing as a learned behavior, not a solver hack.
Continuous Pretraining
Pretrained on a corpus that grows daily, the model learns new parts and packages as the industry ships them — never frozen at a cutoff.
The largest dedicated electronics knowledge base ever assembled.
Chat models know a little about everything. GerberGPT knows everything about one thing — electronics. Parts, packages, parameters, pad stacks. Indexed, vectorized, and continuously updated.
We measured. Twice.
Then a third time.
// 10,000 part-lookup queries across 47 datasheet families. Verified against authoritative manufacturer specs. ChatGPT-5 answered with confidence on every question; only 37.2% were correct.
- ▸ STM32F407VGT6 pinout◉ PASS✕ FAIL · hallucinated PA9 as USART2
- ▸ TPS62840 EN logic level◉ PASS✕ FAIL · cites datasheet rev. 0 (obsolete)
- ▸ 0402 0.1µF X7R replacement◉ PASS✕ PARTIAL · ignores derating
- ▸ LMR16030 thermal pad◉ PASS✕ FAIL · skips PowerPAD wiring
- ▸ ADP1740 dropout @ 1.8V◉ PASS✕ FAIL · gives generic LDO range
- ▸ BAS70-04 forward voltage◉ PASS✕ FAIL · confuses with 1N4148
Find by intent,
not by part number.
Every component parameter — voltage, current, package, tolerance, dropout, ripple, ESR — is embedded into a 768-dimension vector space. Search semantically. Replace BOM items in milliseconds. Find alternates your distributor still has in stock.
- ◉Cosine similarity over 20,000,000+ parameter vectors
- ◉Natural language — engineering English in, parts out
- ◉Distributor stock & lifecycle joined at query time
- ◉Sub-50ms p95 latency, deterministic ranking
// demo · live query latency p50 = 31ms · p99 = 78ms
One revolution at the gate:
every pad knows
its physics.
Pad shape, paste mask, courtyard, thermal relief, current-handling, assembly process — each footprint carries the data the AI needs to place, route, and verify it. This is what makes AI PCB layout tracing actually work.
Stop prompting.
Start engineering /
The schematic library grows every hour. Your next part, footprint, and replacement is already indexed — waiting in the corpus.
- Schematic inference
- Instant component replacement
- Generative footprint & layout
- A model that learns daily