I build ML systems and make them survive production. Lately that means Kubernetes reliability: predicting failures, finding root causes, and proving the model actually works on a real cluster.
I care about whether a result holds up. My repos ship their limitations next to their numbers.
| Repo | What it is |
|---|---|
| otel-sre-copilot | LangGraph agent that investigates Kubernetes incidents through OpenTelemetry (Prometheus, Tempo, Loki). Ships with a reproducible eval harness against 3 baselines, with Chaos Mesh as ground truth. The full benchmark run is still pending. |
| matrix_simple | STA, spatio-temporal GNN for microservice root cause analysis. Selects k suspect nodes first, which drops the cost from O(N²) to O(N+k²). Latency stays flat as the graph grows. |
| ewat | Early warning and anomaly typing on K8s microservices. Separates benign drift from real anomalies, then learns a fault ontology. Validated on a live 9-node cluster. |
| slcmca | slacheck. Answers "what SLA can I actually commit to?" for a multi-cloud microservice app, by walking the Datadog APM call graph and finding the critical path. Catches SLOs you promised but cannot hit. |
| mlops2 | End-to-end MLOps pipeline: scikit-learn to FastAPI to Docker to a VM, deployed by GitHub Actions. |
| Repo | What it is |
|---|---|
| memento | Voice-first memory assistant for Alzheimer patients. RAG plus a knowledge graph, no buttons and no screen, because the interface is the part patients cannot use. |
| voicecall | Voice agent that runs entirely on your machine: faster-whisper, a local Ollama model, Edge TTS, OpenVoice cloning. No API call ever leaves the box. |
| arcad | PREDI-Care, multi-agent clinical decision support for rectal cancer strategy. Built at the A.R.CA.D hackathon. |
| Repo | What it is |
|---|---|
| sklearn2c | Transpiles a trained scikit-learn model into dependency-free C, so it runs on an Arduino or an STM32. |
| abs_arg_dung | Dung's abstract argumentation framework, all 7 semantics. Symbolic AI: given who attacks whom, graph theory decides which arguments survive. |
| prog_diff | An autodiff engine written from scratch in NumPy, up to Conv2d, BatchNorm and Adam. The way to learn what PyTorch does is to rebuild it. |
| ct_visualizer | 2D/3D CT viewer with segmentation overlays, built as a data-quality tool for the PINKCC ovarian cancer challenge, where our team finished 7th of 42. |
Contributor to open source AI infrastructure: a merged PR in litellm, plus work on dspy and opentelemetry-rust.
Also worth a look: a CUDA inference engine written from scratch that beats PyTorch on CNNs by roughly 4x, a 42sh POSIX shell in C, and research on whether a backdoor can survive knowledge distillation into a clean student model. Some of that lives in team or school repos.
Looking for a first full-time ML engineering role in Paris, starting October 2026.