📍 Toulouse, France · 4 years of experience CNES/CDPP
For the past four years I've owned the C++ core and the full-stack pipeline of AMDA, the CNES/CDPP plasma data access service — used by hundreds of researchers and students. Alongside that, I build AI agents and the tools to evaluate them, sitting at the intersection of plasma physics and agentic AI.
Background in astrophysics, space science and planetology (M2, Université Paul Sabatier) — I'm now looking to build agentic systems where the data is complex and the stakes are real.
HelioAI — AI agent specialized in plasma physics
pip install helioai-agent · Open source (MIT)
Answers natural-language plasma physics questions: resolved, fetched, and analyzed in a sandbox.
- Hybrid RAG — BM25 + embeddings (MiniLM) fused via Reciprocal Rank Fusion, over 83k parameters and 70+ missions (AMDA/CDPP, NASA CDAWeb, CSA), indexed in ChromaDB
- Multi-agent orchestration — async streaming agent loop, 4 sub-agents and 6 markdown skills, isolated Python execution (bubblewrap)
- 4 interfaces — CLI, Jupyter, Web UI (FastAPI + SSE), MCP server (stdio and streamable HTTP, Claude Desktop compatible)
- 797 tests, 80% coverage, GitHub Actions CI
HelioBench — Scientific agent evaluation harness
Open source (MIT)
- Three tiers: retrieval across ~83,000 products, plasma physics checked against PlasmaPy, enforced methodology on real satellite data
- No LLM judge — the harness executes and reports; every verdict comes from deterministic code
- Process metrics: number provenance, fabricated identifiers, cost per task,
pass^kover 47 tasks × 3 repetitions
MCP server, RAG, and orchestrator driving sub-agents described in SKILL.md. Write-connected to the AMDA user workspace: derived parameters and reusable time-tables inside the tool.
| AI & Agents | Hybrid RAG (BM25 + dense, RRF), MCP (stdio/HTTP), multi-agent orchestration, sub-agents & skills, tool calling, sandboxing, context engineering |
| LLM & libraries | Groq, Gemini, Azure OpenAI, Ollama, sentence-transformers, ChromaDB, rank-bm25 |
| Languages | C++ (expert), Python (expert), PHP, JavaScript, C, Fortran |
| Full-stack | ExtJS, FastAPI, SSE, SQLite, PHP, REST |
| CI/CD & quality | Jenkins, SonarQube, GitHub Actions, Docker, Git, pytest, ruff, Atlassian |
| Data science | NumPy, SciPy, Matplotlib, Astropy, speasy, PlasmaPy, Jupyter |
- Akkodis — CNES/CDPP · Permanent contract, consultant — 06/2022 → present Lead developer on AMDA: C++ compute core, PHP integration layer, ExtJS front-end. Maintenance, evolutions, CI/CD industrialization.
- CNRS / IRAP · Fixed-term contract — 01–05/2022 Analysis methods for observational forecasts, Euclid mission.
- IRAP — Astrophysics and Planetology · M2 final internship — 2021 Cosmological constraint on the variation of G — publication arXiv:2407.15553
- LESIA — Paris Observatory · M1 final internship — 2020 MHD simulation of the impact of a plasma vortex on Mercury's magnetosphere.
Turkish (native) · French (bilingual) · English (fluent, C1) · Spanish (basic)
📫 furkan.erdogan.pro@gmail.com · Toulouse, France