Senior IA Inference Engineer (x/f/m)

Academic Key; Paris; Hybrid; EUR 90,000 - 140,000; Full time

Job summary

Doctolib is seeking a Senior Inference Engineer to power its AI-driven healthcare products. You will own the self-hosted LLM serving platform, deploy and tune open-weight models, and scale inference across GPU fleets with Ray Serve / Anyscale and Kubernetes-based orchestration.

Responsibilities include managing the model registry, benchmarking, and cost-control while ensuring security and compliance in healthcare data handling across Europe.

Qualifications

• 5+ years of software or computer engineering experience with production LLM serving.
• Hands-on with at least one LLM serving engine (vLLM, SGLang, TensorRT-LLM, TGI, or Triton).
• Proficiency in Python and Terraform; understanding transformer inference mechanics.

Responsibilities

• Build shared LLM serving foundations with MLOps engineers; define serving stack and deployment automation.
• Own model registry, versioning, validation, and promotion to production with quality benchmarks.
• Optimize inference performance, operation, and cost across GPU fleets; ensure reliability and observability.
• Maintain security and compliance in data residency and auditability across environments.

Skills

Python

English fluency

Communication

Tools

Terraform

Kubernetes

Ray Serve / Anyscale

vLLM

TensorRT-LLM

Triton Inference Server

MLflow

Datadog

KServe

Job description

Paris, Paris✓ Link validatedSet a new pulse for healthcare! We are looking for a Senior Inference Engineer to join the ML Platform (TIGER) team in AI & Machine Learning . Your mission will be to power Doctolib's AI-driven healthcare products — from Consultation Assistant to Phone Assistant and Health Companion — by owning the self-hosted LLM serving platform that keeps the most sensitive patient data within infrastructure Doctolib controls. You will build and operate the shared foundations that every AI feature team relies on: deploying, tuning, and scaling open-weight models across our GPU fleet, contributing directly to faster, safer, and more cost-efficient AI for care teams and patients across Europe. Working in the tech team at Doctolib means building innovative products and features to improve the daily lives of care teams and patients. What you'll do Your responsibilities include but are not limited to: Build the shared LLM serving foundations with MLOps engineers — define the serving stack, GPU capacity planning, and deployment automation for open-weight models (Gemma, Mistral, Llama, Qwen, and their fine-tuned variants) on Ray Serve / Anyscale, while shaping the evolution toward Kubernetes-based orchestration Own the model registry and promotion path — establish clear processes for versioning, validating, and promoting models into production, benchmark performance and output quality under realistic workloads, and maintain the curated model catalog together with the Agents Platform team Optimize inference performance and efficiency — operate and tune serving engines (vLLM, Triton Inference Server, TensorRT-LLM) to reduce time-to-first-token, inter-token latency, and cost per token through batching, caching, quantization, and parallelism strategies across our GPU fleet (L4 to H200) Own production reliability and disaster recovery — define SLOs, run load tests before launches, implement safe rollouts, rollbacks, and autoscaling, and participate in on-call and incident response Make serving performance and costs measurable — build Datadog dashboards, alerting, and cost reporting to track latency, GPU utilization, and inference spend, and identify opportunities to improve resource efficiency Keep the inference stack secure, compliant, and evolving — work with Security and Privacy on network isolation, data residency, and auditability, while continuously assessing new models, serving engines (such as SGLang), and GPU generations Who you are Before you read on: if you don't have the exact profile described below, but you feel this job description matches your skill set, we still encourage you to apply. You'll be a great fit if you: Have 5+ years of software or computer engineering experience, including hands-on production experience with at least one LLM serving engine (vLLM, SGLang, TensorRT-LLM, TGI, or Triton) — ideally having chosen, introduced, tuned, or scaled one in production, not just run it with defaults Are proficient in Python and Terraform, with a solid understanding of transformer inference mechanics: prefill vs. decode, KV cache, batching strategies, and why a model is compute- or memory-bound on a given GPU Have strong experience running GPU workloads on Kubernetes or a managed Ray / Anyscale platform, and a performance engineering mindset — you design reproducible benchmarks, read profiler output, and back decisions with numbers Are comfortable owning production reliability: observability, incident response, and on-call are a natural part of the role Have excellent communication and collaboration skills, high agency, a pragmatic build-vs-buy approach, and are comfortable working in a scope that is still taking shape Are fluent in English It would be fantastic if you: Have experience with distributed serving frameworks (KServe, llm-d, NVIDIA Dynamo), including disaggregated prefill/decode or KV-cache-aware routing, or with serving Mixture-of-Experts, multimodal models, or many LoRA adapters on shared base models Have GPU kernel experience or deep profiling skills (CUDA, Triton, Nsight), or have contributed to open-source inference projects Have experience in regulated environments (healthcare, finance, public sector) and with frameworks such as GDPR, HDS, ISO 27001, or SOC 2 Our stack Programming languages: Python, plus comfort reading/debugging a systems language (C++, Rust, or Go) Cloud providers: GCP / AWS / Azure Serving orchestration: Ray Serve / Anyscale (GCP), with Kubernetes (GKE) + GPU scheduling as an emerging parallel target LLM serving engines (in production): vLLM, Triton Inference Server, TensorRT-LLM (also evaluating SGLang and other emerging engines) Models: Self-hosted open-weight LLMs (Gemma, Mistral, Llama, Qwen), fine-tuned in-house Experiment tracking & model registry: MLflow GenAI gateway: LiteLLM LLM evaluation & observability: Braintrust GPUs: NVIDIA L4, A10G, L40S, RTX 6000 Pro, A100 (40GB/80GB), H200 Container / Infra as code: Docker, Terraform, Helm, Kustomize Observability: Dat