ML Engineer – RL & LLM Agent Systems

jda; Paris; Sur place; EUR 75 000 - 115 000; Plein temps

Résumé du poste

The AI Studio is building AI decision-making systems for autonomous supply chains; we’re looking for an ambitious ML Engineer focused on LLMs, agents, and reinforcement learning to help build the training, evaluation, and tooling systems behind robust AI products.

You’ll design experiments, ship production code, improve model behaviour, and create data pipelines and tooling to learn quickly from automated and human feedback, shaping how we use LLMs in agentic systems.

Qualifications

• Experience training or fine-tuning language models (LLMs).
• Proficiency in Python and PyTorch.
• Hands-on with reinforcement learning concepts (RLHF, reward shaping).
• Ability to design and run ML experiments in production.
• Experience building AI workflows and tooling.

Responsabilités

• Design and implement LLM-powered agent environments for decision-making.
• Fine-tune and evaluate LLMs for domain-specific reasoning and decision support.
• Design, test, and iterate on reward functions for LLM agents.
• Review LLM traces and rollouts to understand reasoning and failure modes.
• Identify when reward signals or training setups need adjustment.
• Improve reward models, prompts, and data pipelines based on observed behaviour.
• Build evaluation frameworks for model quality and robustness.
• Create data pipelines for training and human feedback collection.
• Develop tooling to accelerate AI-assisted workflows.
• Experiment with RL, RLHF, RLAIF, and policy optimization.
• Document learnings to guide future work.

Connaissances

Python

PyTorch

RL techniques

LLM training

Experiment design

Description du poste

The AI Studio is building AI decision-making systems for autonomous supply chains; we’re looking for an ambitious ML Engineer focused on LLMs, agents, and reinforcement learning to help build the training, evaluation, and tooling systems behind robust AI products.

You’ll design experiments, ship production code, improve model behaviour, and create data pipelines and tooling to learn quickly from automated and human feedback, shaping how we use LLMs in agentic systems.