A personal AI infrastructure project built on a high-memory AMD platform running fully local LLMs, orchestrated by a lightweight agentic framework and controlled via Discord.
I am a French BTS student specialising in electrotechnics (Électrotechnique), working toward becoming a complete IT/OT engineer, bridging the gap between industrial systems and modern software infrastructure.
This lab is my hands-on learning environment. Beyond the BTS curriculum, I am actively developing skills in:
- Systems architecture — Designing multi-agent AI pipelines, memory systems, and security models.
- Software & DevOps — Linux, Docker, git workflows, Python scripting, REST APIs.
- Cybersecurity — CTF competitions on HackTheBox and TryHackMe, prompt injection threat modelling, containment architecture.
- AI/LLM engineering — Local inference, RAG pipelines, quantization, agentic frameworks.
The philosophy behind this project is simple: the most valuable skill in an AI-augmented world is knowing how to think in systems, not just how to use tools. This lab is where I practice that.
A production-grade personal AI agent platform that runs entirely on local hardware, with a frontier cloud model used only for high-level orchestration. The system is designed to serve as:
- A personal productivity agent — Breaking down complex tasks, managing workflows, drafting communications, all with human approval on consequential actions.
- A studying companion — RAG-powered knowledge base over personal documents, books, and research.
- A skills development environment — Real infrastructure to practice architecture, security, and engineering decisions with genuine stakes.
| Principle | Implementation |
|---|---|
| Local-first compute | 96 GB unified memory platform runs 70B+ models natively. |
| Cloud-assisted orchestration | Claude Pro handles goal decomposition only, never execution. |
| Data privacy | No personal data sent to external APIs. All embeddings local. |
| Approval gates | No outbound action executes without explicit operator confirmation. |
| Incremental complexity | Each phase delivers working infrastructure before the next begins. |
Hardware AMD Ryzen AI Max+ 395 · 96 GB Unified Memory · Fedora Linux
Inference Ollama + ROCm (gfx1151)
Agent runtime OpenClaw (Docker)
Vector DB ChromaDB
Interface Discord (private server)
Remote access Tailscale
Orchestrator Claude Pro (OAuth)
Primary model Qwen3 72B @ Q6 (~55 GB)
Coding model Qwen3-Coder 30B MoE @ Q4
Reasoning model DeepSeek R1 70B @ Q4 / Qwen3 32B @ Q8
Classifier Qwen3 1.7B — routes every request in <100ms
User (Discord)
→ Claude Pro orchestrates & decomposes
→ Lightweight classifier routes each subtask (JSON tag, <100ms)
→ Specialist local model executes
→ Claude evaluates output quality
→ Approval gate (if outbound action)
→ Response delivered
A few non-obvious choices worth explaining:
Why split the classifier from risk assessment? The classifier sees the request, it cannot assess the risk of an action that hasn't been generated yet. Risk tier evaluation (auto-execute / async / blocking) is rule-based and deterministic, applied at the action boundary after the agent has produced its proposed action. Separating these concerns keeps the classifier fast and focused.
Why git for memory monitoring?
Agent memory (MEMORY.md) degrades silently. Bad entries don't crash anything, they slowly misalign behaviour. Weekly git diffs make that drift visible and auditable. Rollback to a known-good memory state is instant.
Why embed before pruning? Pruning without prior ChromaDB ingestion is permanent, unrecoverable data loss. An event-driven tripwire fires before any compaction or pruning event to guarantee mid-session content is embedded first.
🔧 In active development — currently in the hardware verification and base inference phase (Phase 1–2 of the build sequence).
The full architecture document is included in this repository.
- Multi-agent system design with clear separation of concerns.
- Security architecture: containment, least-privilege, defense-in-depth, prompt injection threat modelling.
- Memory system design: tiered persistence, RAG pipeline, git-based monitoring.
- Incremental build methodology with verified phases.
- Hardware-aware model selection and quantization reasoning.
Built as part of a BTS CIEL portfolio to demonstrate self-studying capabilities in software dev, security, systems-thinking and architecture. Designed and documented March 2026.