Descriptif du poste
Location: Paris
Type de contrat : Full-time permanent position (CDI)
Start date: As soon as possible
Compensation: Based on experience and profile
Adlin Science develops a platform for the governance, quality management and exploitation of multimodal biomedical data. We are strengthening the Data team with a senior profile capable of building the platform’s LLM/IA foundation.
The role is not limited to model serving. It covers usage architecture, selection and qualification of open models, local RAG, evaluation, inference optimization, observability and real-world operation on sensitive data
Role positioning
You are the LLM technical referent within the Data team and work closely with the Backend, DevOps, Security, Product and domain expert teams.
You design the components specific to the LLM chain and integrate them into the existing Adlin architecture.
You do not redefine infrastructure, backend or security standards on your own: you rely on the choices, tools and constraints defined with the responsible teams.
You prioritize installable, auditable and maintainable solutions, with no dependency on an external API in production.
Main Responsibilities
Select, test and qualify open models compatible with local deployment and licensing, confidentiality and redistribution constraints.
Define a modular architecture allowing the model, inference engine or RAG strategy to be changed without excessive coupling.
Implement local RAG: embeddings, lexical and vector search, hybrid search, reranking, metadata filters and citation management.
Deploy and operate local inference engines such as vLLM, llama.cpp, TGI, TensorRT-LLM, Triton or equivalent solutions depending on the need
Optimize latency, throughput, memory consumption and stability through quantization, continuous batching, KV-cache management, parallelism and appropriate format choices.
Set up load, resilience and non-regression tests before each production release.
Build pipelines for packaging, versioning, validation, promotion and rollback of models, prompts, embeddings, indexes and configurations
Define a controlled process for importing models and dependencies into the secure environment: signed artifacts, integrity checks, inventory, vulnerability scans and approval procedure.
Containerize components and integrate them with the CI/CD and orchestration tools selected by the Backend / DevOps teams
Produce runbooks, incident procedures, architecture files, test evidence and the elements required for security and quality reviews.