R-002 – Senior Research Engineer, LLM in Paris

Full time, Permanent position Research Paris, France Published on 17 September 2026

About Us

At OVHai LLM, we build large language models and design platforms and services around these models for specific use cases.

We’re a tight-knit team of 15 researchers and engineers, united by curiosity and a shared drive to push the boundaries of what LLMs can do — from training and fine-tuning to rigorous evaluation and benchmarking.

Our team is genuinely diverse and multicultural, bringing together different backgrounds, perspectives, and ways of thinking.

That mix is core to how we work: we believe the best research comes from teams that challenge each other’s assumptions and learn from one another every day.

We take our work seriously, but not ourselves.

Expect a collaborative, low-ego environment, plenty of interesting debates, shared lunches, and lots of fun — alongside real opportunities to contribute to cutting-edge LLM research from day one.

The role

This is a senior-level position for someone who has taken part in training LLMs end-to-end and wants to play a leading role in how our future models are built.

You will work alongside the rest of the research team and is expected to have deep hands‑on experience with at least one stage of large-scale training, strong opinions about what makes a training run succeed or fail, and the ability to turn a paper into a working experiment on a multi-node cluster.

Depending on your background and where the roadmap needs you, you might focus on one of the following: the core pretraining recipe, post‑training and reinforcement learning, data, evaluation, or extending the models to new modalities.

What you will do

• Design, launch and monitor training runs on multi-node GPU clusters (Euro HPC systems and OVHcloud infrastructure), from architecture experiments to full‑scale runs of several billion parameters on trillions of tokens.
• Improve and maintain our training framework: parallelism strategies, memory, throughput, checkpointing, stability and reproducibility.
• Contribute to the recipe at every stage: pretraining data mixture and curriculum, mid‑training, SFT, preference optimisation and RL with verifiable rewards.
• Build and extend the data pipelines that feed these runs: sourcing, filtering, deduplication, synthetic data generation, quality scoring, tokenisation.
• Design evaluations that actually predict downstream usefulness, including domain‑specific benchmarks, and make sure they run continuously.
• Work on our current in‑house LLM architectures and design future ones.
• Mentor junior engineers and interns.
• What we look for
• Several years of hands‑on experience training or fine‑tuning large neural networks, including at least one project on models above ~3B parameters.
• Deep fluency in Py Torch and the modern LLM stack (Hugging Face ecosystem, FSDP/Deep Speed/Megatron‑style).
• Comfortable debugging the full stack: data loaders, NCCL and networking, numerical instabilities, loss spikes.
• Strong software engineering habits: tests, clean abstractions, code review, reproducible experiments.
• Ability to read a recent paper, judge whether it matters for us, and implement it.
• Fluent in both English and French (written and spoken).
• Nice to have
• Post‑training experience: SFT/DPO/GRPO‑style RL, reward modelling, RL environments and verifiers.
• Pretraining data work at scale: web crawls, dedup, quality classifiers, tokeniser training, data mixing.
• Low‑precision training and inference and hands‑on experience with FP8.
• Evaluation and benchmarking infrastructure (lm‑eval‑harness or in‑house), including judge‑based evals.
• A good understanding of Hybrid or state‑space architectures (Mamba, GDN, Hymba‑style designs) and long‑context methods.
• Inference and serving (v LLM, SGLang, Tensor RT‑LLM, custom kernels, Triton/CUDA).
• Experience with multimodal or document‑understanding models.
• Slurm and Kubernetes; experience on HPC centres (Euro HPC, GENCI/Jean Zay, etc.).
• Publications at top venues, open‑source contributions, or public model releases.
• The Recruitment Process
• A 15‑minute introductory call with a senior researcher to discuss the role
• An interview and technical use case assessment with two members of our research team
• An HR interview to assess soft skills
• Debriefing session
• #J-18808-Ljbffr