Org focus (2026-05). This organization is the home of the EU-sovereign LLM serving stack (
ailiance), the autonomous coding agent (agent-kiki), the EDA pipeline engine (euclideda), and the hardware co-design framework (hardware-coagent), plus the MLX fine-tuning toolchain.The FineFab manufacturing platform (
life-*,makelife-*,finefab-*, Factory 4 Life monorepo) was moved on 2026-04-22 to the personal user account github.com/electron-rare for simpler ownership during the pre-launch phase of the L'Electron Rare business entity. GitHub keeps 301 redirects for 90 days; after ~2026-07-22, update remaining clones withgit remote set-url origin git@github.com:electron-rare/<repo>.git.
The active stack — EU-sovereign LLM serving, autonomous coding agent, EDA pipeline engine, hardware co-design, and the MLX fine-tuning toolchain:
| Repo | Vis. | Role |
|---|---|---|
ailiance |
private | EU-sovereign multi-model LLM gateway — Apertus 70B + Devstral 24B + EuroLLM 22B + Qwen3-Next 80B + Granite 4.1 + Ministral, EU AI Act Art. 52/53 transparency, JSONL audit trace |
ailiance-demo |
public | Frontend cockpit for the LLM fleet — public showcase + Tailscale-only admin (FastAPI + Vite + React 19) |
agent-kiki |
public | EU-sovereign autonomous coding agent — VS Code extension + Ink CLI (aki). Fork of Dirac/Cline, routed by default to the eu-kiki LLM gateway, with EU AI Act-compliant JSONL tracing |
euclideda |
private | Open-source pipeline orchestration engine for sovereign EU EDA workflows. Phase 0+1+2+3 closed (5 crates, Runner trait, 61+ tests) |
hardware-coagent |
private | Multi-agent hardware co-design — shared parametric constraints across mecha, PCB, firmware, cost |
mlx |
public | Fork of Apple's MLX — metal-Nx-buffer-limit branches (resource-limit handles + cache-limit bytes patches for large-model training on M3 Ultra) |
KIKI-Mac_tunner |
public | MLX fine-tuning toolkit for Mac Studio M3 Ultra 512 GB — distills Claude Opus reasoning into Mistral Large 123B |
KIKI-models-tuning |
public | FineFab fine-tuning pipeline — model training, evaluation, registry (Unsloth, LoRA) |
kiki-forge |
private | Multi-compute LLM training pipeline for Mac Studio M3 Ultra (post-pivot toward llama.cpp + Qwen3.5 scorer) |
agent-kiki-py-archive |
private | Archived — superseded by the TypeScript agent-kiki |
moodle |
public | Fork of Moodle |
"The boundary between physical and non-physical is very imprecise for us." — Donna Haraway, A Cyborg Manifesto
Bridging embedded hardware and AI — local-first, multi-machine, no cloud lock-in.
| Date | Project | Summary |
|---|---|---|
| 17/04 | clemsail/micro-kiki-v3 |
Cognitive LLM stack for embedded engineering — 35 domain LoRAs on Qwen3.5-35B-A3B, router + negotiator + anti-bias + Aeon memory. Apache 2.0, GGUF, 262K context. |
| 16/04 | L-electron-Rare/KIKI-Mac_tunner |
MLX fine-tuning toolkit for Mac Studio — distills Claude Opus reasoning into Mistral Large 123B. Apache 2.0. |
Training pipeline (KIKI-Mac_tunner) and its output model (micro-kiki-v3) are both open-source. Benchmarks, forks, and negative findings welcome on the discussion thread.
We build monstrous systems — in Haraway's sense. Hybrid organisms where ESP32 firmware and language models share the same nervous system, where a PCB design agent and a fine-tuned LLM are parts of one body.
AI should run where the work happens: on the edge, across heterogeneous machines, orchestrated as a single organism. L'Electron Rare builds the stack that connects embedded firmware to LLM inference to real-time cockpits — without sending a single byte to someone else's cloud.
Every layer is designed to be owned, modified, and deployed by the people who use it.
The system is now organised around three live layers — EU-sovereign LLM serving, the FineFab manufacturing platform, and the autonomous coding agent — all wired to the same fleet of edge nodes:
flowchart TB
subgraph serving["EU-sovereign LLM serving"]
AIL["<b>ailiance</b><br/>Gateway :9300<br/>Apertus 70B · Devstral 24B<br/>EuroLLM 22B · Qwen3-Next 80B<br/>Granite 4.1 · Ministral"]
DEMO["<b>ailiance-demo</b><br/>Cockpit · FastAPI + React 19"]
AK["<b>agent-kiki</b><br/>VS Code ext + aki CLI<br/>EU AI Act JSONL trace"]
end
subgraph finefab["FineFab Platform · github.com/electron-rare"]
direction LR
subgraph platform["Platform"]
LC["life-core<br/>AI backend"]
LW["life-web<br/>Cockpit UI"]
LR2["life-reborn<br/>API gateway"]
LS["life-spec<br/>Spec pipeline"]
end
subgraph hardware["Hardware"]
MH["makelife-hard<br/>KiCad · PCB"]
MF["makelife-firmware<br/>ESP32 · STM32"]
MC["makelife-cad<br/>CAD/EDA"]
end
subgraph eda["EDA orchestration"]
EU["<b>euclideda</b><br/>Pipeline engine<br/>5 crates · Runner trait"]
HCA["<b>hardware-coagent</b><br/>Mecha+PCB+FW+cost"]
end
subgraph ml["ML & Infra"]
KK["KIKI-models-tuning"]
KM["KIKI-Mac_tunner"]
KF["kiki-forge"]
MLX["mlx (fork)"]
FS["finefab-shared"]
FL["finefab-life"]
end
end
AK -- "routes prompts" --> AIL
AIL -- "inference" --> LC
DEMO -. "observes" .-> AIL
EU -. "calls fab/sim tools" .-> MC
HCA -. "constraints" .-> MH
HCA -. "constraints" .-> MF
KM -. "MLX checkpoints" .-> KK
KK -. "GGUF / LoRA" .-> AIL
MLX -. "patched kernel" .-> KM
The previous Agentic Core (mascarade / Kill_LIFE / crazy_life) is now
archived and has been superseded by the FineFab monorepo and the
ailiance + agent-kiki + euclideda stack — see
project_mascarade_deprecated for the
migration notes.
| Project | What it does |
|---|---|
| ailiance | EU-sovereign LLM gateway — multi-model serving (Apertus/Devstral/EuroLLM/Qwen3-Next/Granite/Ministral), JSONL audit trace, alias rewrites, Tailscale-only admin |
| agent-kiki | Autonomous coding agent (VS Code + aki CLI), Cline fork routed to eu-kiki |
| euclideda | EU-sovereign EDA pipeline engine — Rust, 5 crates, dispatch over process/http/MCP |
| hardware-coagent | Multi-agent hardware co-design with shared parametric constraints |
| factory-4-life | Consolidated FineFab monorepo — 21 submodules pinned, single docker compose |
| le-mystere-professeur-zacus | AI-powered escape room: ESP32-S3 firmware + React game engine + real-time voice |
| KiC-AI | AI-powered PCB design assistant for KiCad — local LLM, schematic review, PCB analysis |
| prima-cpp | Distributed LLM inference engine, pipelined-ring parallelism (CUDA + ZMQ) |
| openDIAW.be | AI-powered instruments for live performance — 9 instruments, real-time synthesis |
| ai-novel-engine | Local-first writing atelier — AI generation via Mascarade / Mistral / OpenAI |
Open frontier work on cognition, self-organization, and fine-tuning lives in the sister org hypneum-lab (since 2026-04-20 brand pivot — "GENIAL" stays as the framework acronym used in papers, but the lab is now Hypneum Lab):
| Project | What it explores |
|---|---|
| dream-of-kiki | Substrate-agnostic formal framework for dream-based knowledge consolidation (Paper 1) |
| nerve-wml | Plasticity / GammaTheta multiplexer reference implementation (PyPI v1.8.x) |
| bouba_sens | Benchmarks for sensorial cross-modal binding (TMLR submission track) |
| kiki-flow-research | Wasserstein-gradient-flow engine for micro-kiki self-organization |
| micro-kiki | 32 domain experts (MoE-LoRA) on Qwen3.5-4B base — fits RTX 4090 24GB |
| iact-bench | EU AI Act audit-grade benchmark — 31 domains × ≤23 models, 23 sandboxed Docker validators |
The FineFab platform now lives entirely under
github.com/electron-rare. The naming
reflects the project's evolution:
life-*— core platform services (the "Life" in Factory 4 Life)makelife-*— hardware, firmware, and CAD layers (the "Make" in MakeLife)finefab-*— shared infrastructure and integration (the unified FineFab identity)KIKI-*— ML / fine-tuning pipeline (kept in this org)
| Module | Layer | Role |
|---|---|---|
| life-core | Platform | AI backend — LLM router, RAG, caching, orchestration |
| life-web | Platform | Operator cockpit — Vite + React 19, real-time monitoring |
| life-reborn | Platform | API gateway — Hono, auth, rate limiting, OpenAPI |
| life-spec | Platform | Spec-first pipeline — specifications, BMAD gates, evidence |
| makelife-hard | Hardware | KiCad projects, PCB exports, MCP servers |
| makelife-firmware | Hardware | ESP32/STM32 firmware, PlatformIO, Unity tests |
| makelife-cad | Hardware | CAD/EDA platform — FastAPI + Next.js 15, AI-assisted design |
| KIKI-models-tuning | ML | Fine-tuning pipeline — model training, evaluation, registry |
| KIKI-Mac_tunner | ML | MLX fine-tuning toolkit for Apple Silicon |
| finefab-shared | Infra | Shared contracts — JSON Schema, Pydantic, TypeScript types |
| finefab-life | Infra | Integration runtime — Docker Compose, CI/CD, ops cockpit |
All FineFab subprojects above are assembled inside a single monorepo,
factory-4-life, with 21
git submodules pinned to known-good SHAs. One clone, one docker compose, one
CI-green check across the stack:
| Layer | Submodules |
|---|---|
| Backend | life-core, life-reborn, life-web, finefab-shared |
| Hardware | makelife-cad, makelife-firmware, makelife-hard |
| RAG + Web | rag-web, nc-rag-indexer, e2e-tests |
| Runtime | finefab-life (self-contained docker-compose.prod.yml) |
| Meta | life-spec, life-project, agent-factory-cockpit |
Phase 7 in progress — Industrialisation et Intégration Complète: hardened
Traefik + Keycloak OIDC across every service, shared forward-auth cookie on
.saillant.cc, nightly RAG indexer, Grafana + Jaeger + Langfuse dashboards
wired end-to-end. See the monorepo CLAUDE.md for the full map of 24 nested
CLAUDE.md files and per-domain runbooks.
Factory 4 Life monorepo | 21 submodules | Keycloak SSO | Traefik + Cloudflare | MLX + CUDA fine-tuning | EU-sovereign LLM serving | Local-first, no cloud lock-in
"Monsters have always defined the limits of community in Western imaginations." — Donna Haraway