Looking for a full-time data & AI role
Data & AI Engineer · Paris
I build data pipelines and ML systems end to end: retrieval, MLOps, forecasting.I build software that learns from data, and I make sure it keeps working once real people use it.
About
I started in mechanical engineering and moved into data and AI. At SAP in Barcelona, I helped automate the sales reporting and segmented client accounts. At Akabi, on assignment for Accor, I worked on data pipelines with Snowflake, Terraform and an in-house ETL and orchestration tool. In my own projects, I build the parts that keep a model working after the first demo: tracked training, served models, drift monitoring, and retraining that only ships a better model.
I started in mechanical engineering and moved into data and AI. At SAP in Barcelona, I helped automate the weekly sales reports and grouped client accounts by how they buy. At Akabi, on assignment for Accor, I worked on the systems that bring in and check the hotel group’s data every day. In my own projects, I focus on what happens after the first demo: an AI system that still gives good answers months later, notices when it starts to slip, and is only replaced by a version that proves it is better.
Hybrid search, reranking, answers with citations, and search tools that LLM agents can call over MCP.AI that answers from your own documents, shows where each answer comes from, and can be used by other AI tools.
Syro · insight-mcp
Tracked runs, a model registry with champion and challenger, drift monitoring with PSI, and retraining on real labels.Every version of a model is recorded, its answers are watched in real use, and it is retrained only when the new version does better.
Proths · airport-forecasting
Warehouse pipelines on Snowflake, infrastructure as code with Terraform, and time-series models validated walk-forward.Systems that collect and clean company data every day, and models that predict numbers like airport passengers or wind power.
Akabi for Accor · sdwpf · airport-forecasting
If that's the kind of work your team does, I'd like to hear about it.
01 — Work
Each card starts with the problem, then draws the real architecture that fixes it, checked against its repository.Three AI projects. Each one starts from a concrete problem and shows, step by step, how the system fixes it.
LLM & RetrievalAI assistant
ProblemChatbots answer without sources, and guess when they don’t know.
The fixHybrid search and a reranker over your documents: every answer cites [Source N], or says “I can’t find it”.It answers from your own documents, shows the passage behind each answer, and says when it doesn’t know.
Because an answer without a source is just a guess.
Qdrant · BM25 · bge-reranker · Ollama
Case studyAgents & RetrievalTool for AI assistants
ProblemAI agents need sources they can cite, not guesses.
The fixAn MCP server any agent can call: 3 search tools (BM25 + dense, fused with RRF); the agent’s own model writes the cited answer.A search tool that AI assistants plug into, so they quote your documents instead of guessing.
I wanted agents to look things up, not make things up.
FastMCP · BM25 · fastembed · RRF
Case studyMLOpsAI that maintains itself
ProblemWhen data shifts, a model degrades in silence, and a blind retrain can ship a worse one.
The fixA shift in the predicted-class mix (PSI) raises an alert and does not retrain. A drop in live accuracy does. Promotion needs +0.02 macro F1 on a fixed holdout.A change in the mix of articles raises an alert. It retrains only if live accuracy falls, and switches models only if the new one is clearly better.
Retraining is easy. Knowing it got better is the hard part.
scikit-learn · MLflow · FastAPI · Prometheus
Case studyMore projects
Five more, each with its own page.
02 — Experience
From CAD drawings to production pipelines: mechanical engineering first, then data, then AI that runs in production.
Now
Paris · the projects ↑ · what’s next ↓
Building and shipping the projects on this site: a RAG assistant that cites its sources (Syro), an MCP search server (insight-mcp), a model that retrains itself when data drifts (Proths), and three chatbot evaluations in progress.Building the AI projects shown on this site, and three studies that check whether chatbots give correct answers.
Apr – Oct 2025
Accor Procurement team · Paris
Production data pipelines on Snowflake, with quality checks and alerting on every run, built with an in-house ETL and orchestration tool; dev, staging and prod environments in Terraform, deployed through GitLab CI/CD.I worked on the pipelines that bring in and check Accor’s data every day, and on automating how they are set up and released.
Nov 2023 – Mar 2024
Barcelona · international team · professional English
Helped automate the weekly sales reporting, and segmented client accounts with K-means on their purchasing profiles.I helped automate the weekly sales reports, and grouped client accounts by how they buy.
2020 – 2025
Data major, BI & Analytics minor · diploma awarded in 2026 · semesters abroad in Cape Town and BarcelonaEngineering degree, data specialism, awarded in 2026 · semesters abroad in Cape Town and Barcelona
2018 – 2020
DUT GMP · mechanical and production engineering
2018
Bac S, France
03 — Coming next
Three evaluations in progress. Each one checks chatbots against an official French dataset that holds the right answer.Three studies in progress: I check whether chatbots give correct answers, using official French public data as the answer key.
“How did my MP vote?” Does a chatbot give the real vote, or invent one when the MP was absent or broke with their group?
fetch → ingest → truth → sample → run → grade → report
“Is this pesticide authorised in France?” Do chatbots present withdrawn products as authorised, or mix up EU and French law?
fetch → ingest → checks → sample → run → grade → report
Can a chatbot work out housing aid, the prime d’activité, RSA or income tax for a given household?
reference → freshness audit → simulator check → scoring
Same rule for all three: measure, don’t campaign. Every result gets published, including “no difference”.
Same question every time: does the chatbot know, or does it guess?
04 — Tools
From data to a model in production, and back. Click a tool to see where I used it.The tools I use, from raw data to a model that runs and is watched. Click one to see where I used it.
01Collect & move
02Model & track
03Serve
04Monitor
↺ Monitoring feeds back into the model: a mix shift only alerts. A drop in live accuracy triggers a retrain, as in Proths.
Retrieval & LLMs
Ship & automate · under everything
Languages
Click a tool to see where I used it.
05 — Outside work
Tennis first.
Also skiing, volleyball, swimming and sailing.
Roger Federer fan? Big points. The GOAT.
Psst: try his name in the search at the top.
A drink with friends, the best way to end the week.
Total War first, and Paradox games like Crusader Kings III.
Contact
I’m looking for a full-time data or AI role. Email is the fastest way to reach me.