Our client is an AI safety startup building the safety, reliability, and optimisation layer for advanced AI systems. At the heart of the platform are natural-language policies—rules that clearly define what an AI model should and shouldn’t do. These policies are automatically tested, enforced, and continuously improved at scale.

The company has raised over $10M from leading funds as well as founders and senior leaders from major AI labs and technology companies.

Its systems handle hundreds of millions of API calls every month, and the team fine-tunes and trains proprietary LLMs that run faster and more cost-effectively than both open-source and commercial alternatives.

What You’ll Do

• Train LLMs and VLMs using distributed training frameworks (e.g., Megatron, DeepSpeed).
• Design and implement Mixture-of-Experts architectures for high-efficiency scaling.
• Build multimodal training pipelines handling text, images, and audio.
• Develop custom Triton kernels to optimise training bottlenecks.
• Explore new architectures, hyperparameter strategies, and dataset compositions.

You’re a Great Fit If You

• Have experience fine-tuning large models.
• Understand distributed training deeply: tensor parallelism, pipeline parallelism, expert parallelism, ZeRO, FSDP.
• Can write Triton kernels to accelerate attention, MLP layers, or custom operators.
• Have worked with multimodal systems (e.g., Llava-style, Flamingo-style).
• Care about training stability, convergence, and maximising hardware throughput.

In accordance with local employment laws, applicants must have current, valid authorisation to work in France at the time of application.

If this sounds interesting and you'd like to learn more, click the link below to apply or email me with a copy of your CV on [email protected]

By applying to this role, you understand that we may collect your personal data and store and process it on our systems. For more information, please see our Privacy Notice (https://eu-recruit.com/about-us/privacy-notice/)