AI Research Engineer · agentic systems, LLMs, applied optimization · open to EU roles from September 2026
AI Research Engineer at Orange Labs (Paris), building autonomous AI agents that can independently negotiate, pay for, and activate real infrastructure services from other agents — no human in the loop. My work spans the emerging agentic-web protocol stack (Google A2A, MCP), smart-contract escrow on Foundry, and SDN control of containerized networks. Background: Erasmus Mundus MSc in AI & Optimization (UTC, France); thesis defense September 2026.
Throughline: agent architectures, multi-agent coordination, RAG / LLM systems, and applied optimization .
| Project | What it does | Stack |
|---|---|---|
| ollama-agent-simulation | Thesis PoC — a consumer agent and a provider agent negotiate a bandwidth service, settle payment on-chain for an ERC-721 access credential, and that credential is enforced at runtime by an SDN controller. | Python · FastAPI · MCP · A2A · Foundry/Solidity · Ollama · Docker |
| srl-gnmi-bandwidth-poc | The SDN activation layer: pushes a QoS policer to a Nokia SR Linux PE router over gNMI, enforces it with Linux tc, verifies with iperf3 — all exposed as MCP tools an agent can call directly. |
Python · gNMI/pygnmi · Containerlab · Nokia SR Linux · MCP |
| multi-agent-mcp-demo | Two Claude agents coordinate a buy/sell transaction through a shared MCP server — no orchestrator, pure message-passing, no human after startup. | Python · MCP (FastMCP) · SSE |
| telemetry-rag | Local-first RAG over network-telemetry docs (YANG modules, gNMI captures, snapshots) — Qdrant vector store, Ollama or OpenAI backends, a small index / ask CLI. |
Python · Qdrant · Ollama · OpenAI |
| Instruction-Preference-FineTuning-Pipeline | Reproducible Hydra pipeline for LLM supervised fine-tuning (LoRA / QLoRA) and preference alignment (DPO / ORPO), wired for DeepSpeed/FSDP, W&B/MLflow and LightEval / lm-eval. | Python · PyTorch · TRL · PEFT · Hydra · DeepSpeed |
| claude-usage-widget | Linux desktop widget + terminal CLI showing claude.ai usage limits — GTK3 GUI, truecolor dashboard, local response cache. | Python · GTK3 / PyGObject |
- MSc thesis (defense Sept 2026) — one of the first end-to-end tokenized service exchanges between mutually untrusted AI agents: on-chain settlement for an ERC-721 access credential, enforced at runtime by an SDN controller. → ollama-agent-simulation · srl-gnmi-bandwidth-poc
- Governed multi-agent network automation — separating intent synthesis from execution authority behind an explicit, immutably-recorded approval boundary. → zero-trust-agentic-network-telemetry
- Foundations — working through LLM internals, RAG, and fine-tuning / alignment pipelines from scratch alongside the applied work.
Python · PyTorch · LangChain · Qdrant · Docker · Foundry / Solidity · Linux · gNMI · Containerlab · FastAPI · Git
Full-time Research Engineer / ML Engineer roles in the EU, available September 2026.
- LinkedIn — linkedin.com/in/musel
- Web — musel.dev
- YouTube — youtube.com/@museltabares