AI Platform Engineer at BNP Paribas. I build and run our internal coding agent — a Claude Code style CLI that our engineers use every day, at billions of tokens a month in production. I own the harness behind it: the tool-calling loop, sessions and checkpointing, context engineering and compaction, permissioning and cost control.
Before agents I spent five years on the less glamorous half of this job: MLOps at BNP Paribas, where I took model retraining and deployment from months down to hours, and before that a wind power forecasting system processing terabytes a day of meteorological data.
I also teach Big Data to fourth-year Mathematics students at UNIE.
endstate — agent evals that grade the end state, not the output
Most agent evals grade text: they ask a model whether the answer looks right. That's cheap to build and easy to game. endstate throws away everything the agent said and asserts against what it left behind — did the test suite go green, is there a secret in the diff, did it refuse the destructive command, did it survive being killed halfway through.
The end state is gameable too, which is the half most write-ups skip. So every task pins its test files by hash, refuses new skip markers, confines changes to permitted paths, and runs held-out tests that were never in the sandbox.
22 tasks in disposable Docker containers · deterministic graders that never see the transcript, enforced by signature · a mutation check that removes each guard and proves the suite notices · 344 tests · mypy --strict
pip install endstateDocs · How agents actually work — a walk through harness internals: the loop, the sandbox boundary, permissioning, compaction, checkpoint durability
pysuricata — exploratory data analysis built on streaming algorithms
Profiles a DataFrame in a single pass, with memory that stays bounded no matter how large the data is. Welford/Pébay for exact moments, KMV sketches for distinct counts, Misra-Gries for heavy hitters, reservoir sampling for quantiles. Pandas, Polars and LazyFrames; output is one self-contained HTML file with no external assets.
pip install pysuricataDocs · Live example report · Statistical methods
Big Data course labs — open materials for the course I teach at UNIE
Out-of-core computing and storage formats, streaming and sketching algorithms, distributed systems from first principles, Spark, Airflow, Slurm.
cka-practice — eleven mock CKA exams solved against a real cluster
The grader inspects the cluster's actual state, not an answer key. Same idea as endstate, applied to Kubernetes.
Also, for fun: 2048 — RL agents (DQN, CNN) trained with PyTorch on MPS, Rust core compiled to WebAssembly. And Tetris — pure-Rust engine in WASM with SRS wall kicks, 7-bag randomiser, and AI opponents.
Languages Python · SQL · Bash
Agents & LLM LangGraph · Pydantic-AI · MCP · harness design · context engineering · tool-calling loops · sessions & checkpointing · token budgeting & cost control · evals · RAG
Platform & infra Kubernetes (CKA) · Docker · Terraform · GitOps · GitLab CI/CD · GitHub Actions · Domino Data Lab · JFrog Artifactory · FastAPI · Linux
Data & ML MLflow · PyTorch · TensorFlow · scikit-learn · XGBoost · Spark · Airflow · Dask · Xarray · Kafka · PostgreSQL · S3 · pandas · Polars · NumPy · streaming algorithms
Engineering uv · Pytest · Coverage.py · Ruff · mypy --strict · Pydantic · pre-commit · Conventional Commits · MkDocs · trunk-based development
- MSc Industrial Mathematics (M2i), modelling specialisation — Universidad Carlos III de Madrid
- BSc Mechanical Engineering (bilingual) — Universidad Carlos III de Madrid, with an exchange year at Purdue University
- CKA: Certified Kubernetes Administrator — CNCF
- Deep Learning Specialization & MLOps Engineering for Production — DeepLearning.AI