Henri Bertrand

— Design, deployment, and operation of the BNP group's AI inference platform, providing LLM and ML capabilities to all entities (standardized and custom models).

— Operation of a multi-site on-premise GPU cluster via HyperShift, hosting dedicated AI, HA, and inter-site redundant OpenShift clusters.

— Implementation of OpenShift AI clusters integrating Kubernetes, SDN, Service Mesh, Operators, Prometheus, Grafana, Alertmanager, Loki, Jaeger, Pipelines, RBAC, and Network Policies.

— Sizing of multi-GPU nodes for models from 7B to 600B parameters, MIG optimization, scheduling, NUMA, and NVLink topologies.

— Operation under industrial constraints: tens of thousands of concurrent users, >150k MAU, strict SLAs, optimized TTFT, p99 latency < 3s.

— Advanced scaling, batching, and prioritization strategies on shared non-production clusters and dedicated production clusters.

— Serving of LLMs, embeddings, and financial ML models (scoring, forecasting, anomaly detection) on shared infrastructure and isolated, encrypted production environments.

— Design of strong network, compute, storage, and secrets isolation for sensitive contexts.

— HA NAS hybrid architecture + shared local storage for performance and fault tolerance.

— Multi-site redundancy, DRP, backups, and service continuity.

— Structuring product governance: roles, committees, offer lifecycle, service catalog, and internal contracting.

— Vendor and critical dependency management.

— Operation of the Red Hat ecosystem: OpenShift, OpenShift AI, HyperShift, Quay, ACM, ArgoCD, Pipelines, Service Mesh, Keycloak, ODF.

— Alignment with group standards for security, compliance, observability, and operations.

— Design of advanced RAG agents (ReAct, Multihop, Plan-Search-Respond) for Risk Management, Audit, Business Analysis, and IFRS using Python, Haystack, LangGraph, DSPy, LiteLLM, Pydantic, Azure OpenAI, Mistral.

— Production deployment of a multi-risk report generation agent (climate, geography, human rights) via LangChain, Tavily, GPT-4o, and Llama 3.1.

— Multi-level indexing strategies, peripheral context management, hybrid search (chunk, embeddings, full-text).

— Indexing of images and non-textual content in documents (GPT-4o, YOLO, Azure OCR, ColPali).

— Industrialization of CI/CD for Data Science projects: build, tests, packaging, deployment, and monitoring of ML/LLM pipelines.

— Co-design of the Azure AI foundation with the IT department: Azure ML, AKS, Blob, Functions, and Durable Functions.

— Inference architectures combining streaming, batch, and event-driven orchestration via queues and message buses.

— Distributed asynchronous pipelines (fan-out/fan-in, retry, idempotence, fault tolerance).

— Azure ML model deployment: autoscaling, versioning, blue/green, canary, rollback.

— SOTA evaluation stack: context relevancy/recall, ATS, nDCG@k with dedicated pipelines.

— Setup of agent store, config store, and dataset store for governance.

— Tracking of LLM costs by user/use case with quotas and alerting.

— Technical leadership of a team of 4 Data Scientists.

— Management of DSLP backlog + Scrum in Azure DevOps (KANBAN, boards by use case).

— Creation of a dedicated AI codebase following Python/DS best practices: uv, pre-commit, Makefile, DevContainer, Ruff.

— Comprehensive documentation of algorithms, metrics, and indexing.

— Unit, integration, and E2E testing strategy.

— Code quality: pylint, black, isort, bandit, safety, ruff, mypy, coverage integrated into CI/CD.

— Use case qualification with program management.

— Design and industrialization of a decision support platform for patients with chronic kidney disease, operated in production under health data constraints (security, sovereignty, compliance).

— End-to-end architecture: ingestion, normalization, pseudonymization, RAG engine, LLM stack, inference layer, business API, and user interfaces.

— Multi-source medical RAG engine leveraging patient records, biology, and clinical repositories (FAISS/Qdrant, biomedical embeddings, hybrid retrieval, reranking, longitudinal context management).

— Clinician interface similar to a decision support chat with context visualization, response justification, and feedback (Gradio).

— Product management: roadmap, iterations, user workshops, and impact measurement on decision quality.

— Fine-tuning of Llama-3 8B, Mistral 7B, Qwen on medical corpora (Transformers, PEFT, QLoRA/LoRA, TRL).

— Supervised alignment and RLHF pipelines with human-in-the-loop.

— Comprehensive governance: dataset/model/prompt versioning, metrics, audits, and traceability of clinical decisions.

— Responsibility framework: confidence thresholds, human fallback, controlled refusal, and medico-legal traceability.

— HA bare metal platform based on vLLM (multi-model, continuous batching, KV cache, tensor parallel, GPU scheduling) and Infinity for large-scale embeddings.

— Kubernetes orchestration of AI/data services: API, vector store, PostgreSQL, monitoring, MinIO encrypted storage, CI/CD, and audit logs.

— Operational processes: SLA, technical and business monitoring, incident management, and service continuity.