I'm an AI Engineer at SFEIR, a French IT consulting firm, currently on mission at Decathlon France building DaiLY, a multi-agent HR assistant for 30,000+ employees in Google Chat. A systematic eval pipeline lifted its lead agent's rubric pass rate from ~60% to 97.7%. Before joining SFEIR, I spent five years inside Decathlon: first in France on internal HR tooling, then in the UK building the e-commerce marketplace platform that drove £2.6M GMV in 2024.

I work embedded inside client teams, owning delivery end-to-end. My open-source projects tackle problems I've met running LLMs in production: privacy-safe RAG, cost attribution and quality evaluation.

• 🛠️ Multi-agent systems · RAG · LLMOps · evaluation pipelines
• 🌍 🇫🇷 French (native) · 🇬🇧 English (C2) · 🇪🇸 Spanish

Four open-source projects engineered to production standards: Pyright strict, 86–95% test coverage, CI and eval workflows on every push. Together they cover the three problems every team running LLMs hits:

• Cost: gateway tracks where the budget went; autopilot prevents it going to the wrong place
• Quality: detector catches quality drops when prompts change
• Privacy: guardian keeps personal data out of both the index and the response

An OpenAI-compatible gateway that attributes spend across the four stages of a RAG pipeline (retrieval, reranking, generation, evaluation), so teams stop guessing which stage is eating their budget.

RAG-aware cost attribution · multi-provider fallback · circuit breakers · Redis rate limiting · 92% coverage

A two-stage router (embedding similarity, then DeBERTa zero-shot on ambiguous cases) that sends each request to the cheapest capable model, then an async verifier re-runs a sample on a stronger model and logs every routing disagreement to a correction dataset.

94.6% agreement with a stronger-model verifier · 14.6% cost reduction on a 1,000-request load test · 95% coverage

A CI quality gate that runs your LLM against a golden dataset on every PR, diffs accuracy with Wilson 95% confidence intervals, and blocks the merge when the drop is statistically real.

-30pp regression detected automatically in CI · 86% coverage · GitHub Actions + Slack alerts

A RAG pipeline with three-stage PII detection at ingestion (Presidio, GLiNER, then DeBERTa validation on low-confidence spans only) and a post-generation audit on every answer, designed around EU AI Act data-governance principles (Art. 10).

100% PII recall · 0.93 precision on a 91-entity synthetic benchmark · 0 post-generation leaks · 93% coverage

Lead developer on a multi-agent HR assistant in Google Chat serving 30,000+ employees in France, answering in 5 languages.

• Coordinator + 4 specialized sub-agents over the A2A protocol on Cloud Run, built on Google ADK and Gemini (Vertex AI)
• Vertex AI Search RAG pipeline over 100+ HR policy documents, unifying HR knowledge scattered across 280+ internal tools; every answer passes a grounding-evidence gate before reply
• LLM-as-Judge eval pipeline: 600+ golden cases across 4 agents + a coordinator routing suite; rubric pass lifted from ~60% to 97.7% on HR Knowledge, 87–96% across the remaining agents
• Production security: prompt-injection screening (Google Model Armor), signed agent-to-agent calls, fail-closed PII redaction on LLM traces, EU data residency
• 2-layer production kill switch (5–10s Cloud Run cutoff + 30s TTL registry toggle, no redeploy) · keyless CI/CD via GitHub Actions + Workload Identity Federation
• BigQuery observability tying answer quality to the exact prompt revision (per-prompt-hash, per-model, per-cost-center)
• Technical alignment with Decathlon NL and ES on cross-country RAG architecture

Built the e-commerce marketplace connector platform across three countries.

• 8 Java/Spring Boot microservice connectors across UK, South Korea, and Switzerland
• £2.6M GMV in 2024 · €528K GMV on the Glovo connector since August 2025
• 40,000+ product updates/day via Cloud Firestore
• Onboarding time per new marketplace: 8 weeks → 4 weeks
• Built a cost-free user-management interface in Google Apps Script: 90% reduction in fraudulent transactions across all 90 UK retail stores

Automation of HR processes and internal tooling.

• Built a Java/Spring Boot aggregator integrating with Greenhouse webhooks, cutting manual data entry by 50%
• Streamlined contract generation and internal API workflows

• AI & agents: Gemini · Vertex AI · OpenAI · Claude · Ollama · Hugging Face · Google ADK · A2A protocol
• RAG, evals & safety: Qdrant · MLflow · Databricks · OpenTelemetry · Vertex AI Search · LLM-as-Judge · Presidio · GLiNER · Model Armor
• Backend: Python · FastAPI · Java · Spring Boot · Kafka
• Cloud & data: Google Cloud · BigQuery · Firestore · Terraform · Docker · GitHub Actions · PostgreSQL · Redis · Prometheus · Grafana · Cloud Run