README.md

metadata

title: Ailiance
emoji: 📊
colorFrom: blue
colorTo: yellow
sdk: static
pinned: true
license: apache-2.0
short_description: EU-sovereign AI for hardware design

Ailiance

EU-sovereign AI infrastructure for hardware design. OpenAI-compatible gateway with hardware-specialist LoRA routing. Aligned with the EU AI Act (Art. 52, 53 — GPAI fine-tunes).

🌐 ailiance.fr · 💻 github.com/ailiance · 🎮 Bench Playground

---

What we ship

A production gateway plus a curated library of small, domain-specialist LoRA adapters that beat large generalist models on hardware-design tasks while running on commodity Apple Silicon and EU-hosted GPU.

• Gateway — FastAPI, Jina v3 embeddings + MLP router, 24 OpenAI-compat aliases, routes to the right specialist backend (Mistral-Medium 128B, Gemma-4 + eu-kiki LoRA, Qwen3-Next 80B MoE, Granite-30B, EuroLLM 22B, plus 13 mascarade hardware experts).
• Models — 15 LoRA adapters and routers on this org, all Apache-2.0.
• Datasets — 13 hardware-domain datasets (KiCad, SPICE, STM32, EMC, embedded, IoT, power, DSP, PlatformIO, FreeCAD, kicad9plus corpus, kill-life embedded QA).
• Bench — ailiance/ailiance-bench, 7-task hardware evaluation (KiCad DSL/PCB generation, SPICE reasoning, schematic extraction, ERC analysis). Try the interactive playground.

Phase 6 bench — champions

Base = gemma-e4b-eu-kiki-base. Δ shown in points (× 100) vs base.

Task	Winner	Δ vs base
P1 kicad-dsl	eu-kiki	+55
P1 kicad-pcb	eu-kiki	+42
P1 spice-sim	eu-kiki	+25
P3 kicad-sch-extract	mascarade	+48
P3 kicad-sch-extract	eu-kiki	+38

Verdicts

• 🥇 eu-kiki — generalist champion (4/7 tasks). Hosted at macm1 :8502.
• 🥇 mascarade — P3 extraction champion (+48 pts). Hosted at Tower Ollama :8004.
• ⚠️ kicad9plus — catastrophic forgetting on SPICE/P2/P3. Use only in permissive-KiCad-only contexts.
• 🚫 kicad-sch from-scratch — unresolved across all adapters. Bottleneck: KiCad 6+ S-expr absent from pre-training corpus.

Full scoreboard: ailiance-bench README · commit 46801af.

Featured models

Generalist adapters (Apache-2.0)

• devstral-v3-sft — code-tuned Devstral 24B
• apertus-electronics-hw-lora — Apertus 70B + hardware fine-tune
• eurollm-multilingual-eu-lora — EuroLLM 22B + EU multilingual

Domain-specialist LoRA (Qwen3-4B base, Apache-2.0)

• qwen3-4b-mascarade-kicad-lora
• qwen3-4b-mascarade-spice-lora
• qwen3-4b-mascarade-stm32-lora
• qwen3-4b-mascarade-emc-lora
• qwen3-4b-mascarade-embedded-lora — 🥇 P3 champion
• qwen3-4b-mascarade-iot-lora
• qwen3-4b-mascarade-dsp-lora

Featured datasets

Dataset	License	Purpose
mascarade-{kicad,spice,stm32,emc,embedded,iot,power,dsp,platformio,freecad}-dataset	CC-BY-SA-4.0	Hardware Q&A per domain
kicad9plus-permissive	CC-BY-SA-4.0	KiCad schematic corpus, permissive
kicad9plus-copyleft	GPL-3.0	KiCad schematic corpus, copyleft (upstream lib)
kill-life-embedded-qa	CC-BY-SA-4.0	Embedded systems Q&A

EU AI Act compliance

All Apache-2.0 models on this org are tagged with:

• Art. 52 — transparency obligations for GPAI fine-tunes
• Art. 53 — content provenance + training-data summary
• gpai-fine-tune — declares the model as a fine-tune of a GPAI

Datasets are dual-licensed (permissive vs copyleft) when upstream constraints require it. Provenance and license chain documented per dataset card.

License

• Models: Apache-2.0 (all Apache-2.0 declared on this org)
• Datasets: CC-BY-SA-4.0 except kicad9plus-copyleft which is GPL-3.0
• Code: Apache-2.0 (see github.com/ailiance/ailiance)