I'm Lead AI Engineer at SFEIR, a French IT consulting firm, currently on mission at Decathlon France building DaiLY, Decathlon's multi-agent assistant for internal HR, Workday and training procedures. It runs in Google Chat, is available to 30,000+ employees and answers in 5 languages. A systematic eval pipeline lifted its lead agent's rubric pass rate from about 60% to over 97%. Before joining SFEIR, I spent more than four 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: cost and routing, quality evaluation, privacy-safe RAG and multi-agent controls.

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

Six repos that work as one system for running LLMs in production: meter and route every call, keep personal data out, run agents with production controls, then turn their traces into eval cases that block a bad merge. Each repo keeps a written log of its design decisions.

An OpenAI-compatible gateway that splits LLM spend across the four stages of a RAG pipeline (retrieval, reranking, generation, evaluation), with provider fallback, circuit breakers and Redis rate limits.

2.14 ms p50 gateway overhead per request; one worker holds 250 req/s at p99 6 ms · design decisions

A router the gateway can call to send each request to the cheapest capable model, with a background verifier that checks the cheap answers.

1.3% cost reduction on 1,000 requests, 500 of them real NQ-Open questions (simulated pricing) · design decisions

A CI quality gate that scores your LLM on a golden dataset on every pull request and blocks the merge when the drop is statistically real, using Wilson 95% confidence intervals.

Flags a -30 pp drop (86.8% to 56.6%) as critical on the incident-triage demo (mock model) · design decisions

A RAG pipeline that redacts PII at ingestion, screens queries and documents for prompt injection, and audits every answer before it reaches the user.

100% PII recall, 0.93 precision on a 91-entity synthetic corpus · design decisions

A shopping and after-sales assistant built as a small multi-agent system (LangGraph coordinator, Pydantic AI specialists, A2A and MCP), with signed agent cards, human approval before any write and a kill switch per agent.

design decisions

Turns traces from the gateway and agent-mesh into reviewed, PII-free golden cases that llm-regression-detector runs in CI.

design decisions

Lead AI Engineer on DaiLY, a multi-agent assistant for internal HR, Workday and training procedures in Google Chat, available to 30,000+ employees in France and answering in 5 languages. Built from scratch, in production since May 2026.

• Coordinator + 4 specialized sub-agents (2 in production, 2 in beta) 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: 800+ golden cases across 4 agents + a coordinator routing suite, 92 to 100% rubric pass per agent; the lead agent went from about 60% to over 97%
• 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