GBM Scorer
Gradient-boosted ensemble trained on 6,800 real and synthetic designs. Scores signal integrity, power design, layout quality, component selection, thermal management, and documentation — each with calibrated confidence.
Now in open beta
AI that scores
real engineering
work.
Upload a KiCad PCB layout, schematic, or SPICE netlist. EngineIQ returns 6 competency scores with theoretical confidence bounds, component-level fault localization, and IPC-2221 grounded fixes — in 250ms.
$ curl -X POST /api/v1/review \
-F "file=@stm32_board.kicad_pcb"
{
"overall_score": 82.4,
"grade": "B",
"srhn_enrichment": {
"pac_lower": 0.643,
"pac_upper": 0.961
},
"localized_faults": [{
"fault_type": "trace_too_narrow",
"fix": "Increase to ≥0.78mm",
"inference_ms": 51.3
}]
}
$
Platform statistics
0.942
Spearman r (accuracy)
51ms
Inference time (CPU)
192
Feature dimensions
17
IPC-2221 fault types
Under the hood
Three systems, one response
Every design review combines a trained gradient-boosted model, a C11 reasoning engine, and IPC-2221 rule-based fault localization — all in under 250ms.
SRHN v4 Reasoning
Semantic Reasoning Hypergraph Network — a 200KB C11 binary with zero dependencies. Provides PAC-Bayes theoretical confidence bounds [R7], causal chains (narrow_trace→overcurrent→thermal), and contradiction detection.
Fault Localizer
17 IPC-2221-validated fault types. Points to the specific component reference (e.g. "C3", "TRACE at 45.2mm,12.8mm"), gives the exact fix, and when known, the correct replacement part number.
Technical depth
Built for engineers, by engineers
⬡ Custom Model Builder
Your org's standards.
Your model. Your moat.
A generic model trained on internet data has no concept of your fab's DFM tolerances, your approved component library, or your senior engineers' judgement. EngineIQ trains a private ML model on your own designs and expert annotations.
- ✓Fine-tuned GBM + optional GATv2 GNN on your corpus
- ✓Private SRHN graph trained on your design guidelines
- ✓Accuracy improves: r=0.942 → r≈0.97+ after 500 annotations
- ✓Impossible to replicate — trained on your non-public data
Custom model pipeline
📁
Your design corpus
.kicad_pcb / .step / SPICE — minimum 50 designs
↓ feature extraction (192-dim)
🧑
Expert annotation
Your engineers annotate 200+ designs (3–5 min each)
↓ GBM retrain + SRHN ingest
🔒
Your private model
Versioned, content-hashed, gated by golden test set
⚡ Optional: GATv2 GNN fine-tuning on your topology patterns (A100 ~8h) pushes r to ~0.98
From the blog
Engineering AI in practice
Why PAC-Bayes bounds make engineering scores defensible
Traditional ML confidence is empirical. PAC-Bayes provides theoretical guarantees derived from spectral norms — the difference between "we think" and "we can prove".
SRHN v4: 8 research modules for engineering reasoning
Multi-semantic orthogonal views, hyperedge attention, and Laplacian spectral reranking — how we built a reasoning engine that actually understands circuit topology.
The behavior bonus: why we score how engineers work, not just what they submit
An engineer who used the trace width calculator before submitting scores higher on signal integrity. This is a genuine, hard-to-fake assessment signal.
Join us
Invest or collaborate
We are seeking seed investors and senior engineering collaborators to help us build the definitive technical assessment platform.
$750K
Seed round target
- GPU compute for GNN pretraining (A100, ~3 days per run)
- 3 senior EE annotators — 200+ expert-labeled designs
- Go-to-market: technical hiring platforms, bootcamps, cert providers
- Enterprise security hardening (SOC2 Type II, multi-tenancy)
- 12-month runway to first paying enterprise customer
For engineering collaborators
We need senior PCB, analog, power, and embedded engineers to annotate designs and validate scoring. Equity + recognition for meaningful contribution.
Contribute on GitHubInvestor Pitch Deck
5-slide overview: problem, solution, market, traction, ask
Ready to assess engineering work with AI?
Upload your first KiCad design or contact us to discuss enterprise deployment. We respond to every email within 24 hours.