Rafik Mammeri

I take AI agents from whiteboard to production — and make them part of how the company works.

Senior AI Engineer and technical lead for AI agents at Boulanger, one of France's largest electronics retailers (€4B). Four systems in production across three channels: customer web & mobile, voice, and the tools employees use every day. Before that, seven years shipping machine learning in regulated banking. PhD in mathematics. Based in Lille, France.

~2,000conversations / day 75%resolved end-to-end 78%satisfaction, 10,000+ ratings 6 mofrom zero to production

Explore the projects → Read the writing → Full experience →

How I work

Architecture before prompts

The biggest latency and cost wins in LLM systems are structural. The assistant answers in one or two model calls per message not because of clever prompting, but because routing and handoffs are native to the graph.

Proof: the no-orchestrator design · the article

Design for the model being wrong

Four independent security layers stand between an agent and the data warehouse. Even a hallucinated query lands on a read-only role scoped to its use case — and gets rejected.

Proof: Self-BI's four-layer model · regulated-banking roots

The business owns the product

The HR assistant is owned by HR — content, scope, response validation. My job is making sure the technical layer never gives them a reason to doubt it.

Proof: the HR product-owner model

Production is a loop, not a launch

Versioned prompts, LLM-judge scoring against live traffic, and continuous releases since day one. Quality in production isn't achieved — it's maintained.

Proof: the quality loop at Boulanger

Skills

LLM & Agents

LangGraphLangChainGoogle ADKMCPRAGAzure OpenAIGeminiLangfusePrompt systems

6 specialized agents, 26 production tools

AI Security & Governance

GuardrailsPrompt-injection defensePer-agent tool isolationSQL enforcementSecrets isolationGDPR-aware tracing

Four independent layers in front of the warehouse

Engineering & MLOps

PythonFastAPIAsync / SSEHexagonal architectureKubernetesDockerCI/CDSonarQube

~2,000 conversations/day without blocking a worker

Cloud & Data

GCP / Vertex AIAzureSnowflakeMongoDBTerraformPrivate networking (PSC)API gatewaysAlgolia

Google Cloud Professional Data Engineer

ML Foundations

XGBoost / CatBoostCredit scoring (Basel II/III)Time series (LSTM)Transformers (SBERT)MLflow

Seven years of regulated, production ML before LLMs