Théo Martin

Tech Lead for Akeneo's Core AI team. Design, deployment, and operation of the internal AI platform used by all product teams for 900+ enterprise retailers.

- Architected and deployed the AI inference platform in production: 30M+ product attributes generated by LLM/month, 99.999% availability, p95 under 150ms

- Designed and delivered the Data Architect Agent (DAA), an agentic system reducing catalog onboarding time from several months to a few days for enterprise retailers

- Led the multi-agent orchestration layer in production (Google ADK + LiteLLM proxy): autonomous catalog enrichment and approval workflows with human gates, JWT/ACL inheritance, and SSE streaming, estimated reduction of 60%+ in manual approval cycles

- Designed zero-code multi-model routing with automatic failover (GPT, Claude, Gemini, Mistral): centralization of all LLM traffic via a single proxy, 40% reduction in inference costs through cascading and caching, 70% reduction in log volume

- Principal technical referent for 5 cross-functional teams (product, engineering, design, marketing, security). Comprehensive AI Governance: drift monitoring, hallucination control, audit trails, personal data guardrails, and compliance

Stack: Python, GCP (Vertex AI, BigQuery, CloudRun), FastAPI, LiteLLM, Google ADK, Docker, Kubernetes, LangChain,

First ML engineer at the startup. Technical work directly contributed to acquisition by Akeneo. Active participation in technical due diligence (architecture, pipelines, ML infra).

- Designed and operated the complete MLOps pipeline on GCP (ingestion, BigQuery, training, serving) for enterprise retail groups (Rexel): 90%+ extraction accuracy, outperforming the client's internal data science team, millions of SKUs processed/month without downtime

- Optimized language models (FLAN-T5, pre-ChatGPT) for product categorization on complex industrial catalogs: 97% accuracy, 40% memory reduction through quantization and custom batching, 10x GPU acceleration, near-perfect zero-shot generalization on new catalog families

- Implemented production observability and automatic retraining loops with drift detection: 85% fewer degradation incidents, 35% reduction in inference costs

Stack: Python, PyTorch, Transformers, HuggingFace, Docker, GCP (BigQuery, Vertex AI)

Master's degree courses in Artificial Intelligence:

- Introduction to Machine Learning and LLM Functionality

- Overview of AI Careers and Business Case Studies