Grégoire Mureau
📘 Ce profil est également disponible en français 🇫🇷
ML / data engineer — RAG · LLM serving · homelab Kubernetes
Graduate AI Engineer (Data Science Expert). Eighteen months in production data engineering at Ifremer (Copernicus Marine, millions of in-situ observations per day; 3 scientific publications).
This Forgejo instance is part of the infra: it is not a third-party host. GitHub (GrgMdmn) is the public mirror; this is day-to-day.
Homelab — compute vs serve
Two Ubuntu Server machines, two independent k3s clusters (not federated: if one is down, the other must keep serving). They talk over Tailscale.
| Mini PC (compute) | NAS (serve) | |
|---|---|---|
| Hardware | AMD Ryzen 7 8745HS · Radeon 780M iGPU · 48 GB unified RAM | Intel N100 · 31 GB · MergerFS + SnapRAID |
| Role | Inference, CI runners, observability | Public edge: Traefik, git, cloud, ML demos |
| Orchestration | k3s (single-node) | k3s (single-node) |
The 780M is a modest iGPU with no dedicated VRAM. A discrete GPU in the same form factor (mini PC / laptop) typically tops out at 8–16 GB of VRAM; here model weights sit in unified memory (GTT, tens of GiB). That is what makes it possible to host and serve an MoE such as Qwen3-30B-A3B, which would not fit in the VRAM of an equivalent dGPU.
Try it (temporary chat, nothing persisted): Qwen3-30B MoE via Open WebUI
Also live from the NAS: ML APIs · Forgejo (this page).
Repositories listed under this user account are mostly homelab / infra. The OpenClassrooms Master lives in the openclassroom-ai organisation (Forgejo does not pin org repos on a user profile).
Case study — sovereign community platform
A community founded in 2007 (19 years, approaching 20), hosted first on one proprietary platform then a second, migrated to open-source Discourse on a dedicated VPS I operate (not on the NAS): depiedencap.org.
- 22,230 topics · 424,835 posts · 7,102 accounts mapped · 93,884 images (8.8 GB)
- Custom Discourse plugins (onboarding, French locale/PWA, RAG citations)
- Mail & GDPR ops: reactivation campaigns, erasure workflow with evidence files
- Production-grade RAG agent (depiedencap-rag): post-level embeddings (bge-m3 → pgvector, ~428k posts), cross-encoder reranking, conversational query rewriting, server-side citation whitelisting, deterministic input gate — measured 0.97 Hit@5 / 0.89 MRR on 812 real questions, 20/20 on safety eval. Guides members to existing threads instead of answering for them.
Scraping pipelines and member data stay private — this page is architecture and outcomes, not a source dump.
Master AI Engineer — OpenClassrooms
Organisation: openclassroom-ai
GitHub mirrors (same names): github.com/GrgMdmn
| # | Project | Forgejo | Live |
|---|---|---|---|
| 2 | Paris trees (Pandas, Folium) | oc-p02-paris-trees | — |
| 3 | OpenFoodFacts (cleaning, PCA, ANOVA) | oc-p03-openfoodfacts | — |
| 4 | Credit scoring (LightGBM, SHAP, LIME) | oc-p04-credit-scoring | — |
| 5 | Olist customer segmentation (RFM, K-means) | oc-p05-customer-segmentation | — |
| 6 | Product classification (NLP + CV) | oc-p06-product-classification | — |
| 7 | Sentiment analysis (LSTM, DistilBERT, FastAPI) | oc-p07-sentiment-api | demo |
| 8 | Street vision (U-Net, FPN) | oc-p08-street-vision | demo |
| 9 | Street vision PoC (SegFormer vs FPN) | oc-p09-street-vision-poc | demo |
| 10 | Content recommendation | oc-p10-content-recommendation | — |
| 11 | Big data fruits (PySpark, EMR) | oc-p11-bigdata-fruits | — |
| 12 | Fashion-Insta project framing | oc-p12-fashion-insta | — |
ML landing: api.gregoiremureau.com/ml/
Links
- Website: gregoiremureau.com
- GitHub: github.com/GrgMdmn
- LinkedIn: gregoire-mureau
- Publications: ResearchGate
- Produce Estimates (React PWA, anonymised demo): demo.gregoiremureau.com/estimaciones_demo/