Applied Science - Time Series

Mistral AI is seeking Applied Scientists and Research Engineers focused on time series. You will design, train, and deploy state-of-the-art time-series foundation models, and apply them to the most demanding forecasting problems in industry, from power grids and plants to trading desks.

Timestral, our family of time-series foundation models, ranked first on amongst all benchmarks at release (tech report: https://arxiv.org/pdf/2610.37585v1 ). The role has two sides: building the next generations of Timestral with our Science team, and taking them into production with our industrial and financial clients, adapted to their data. Working cross-functionally with both external and internal science, engineering, and product teams you will deliver high-impact AI solutions. We are made of researchers, engineers, product builders, ex-founders who love tackling the most complex real world problems in the industry.

What you will do

• Run pre-training, post-training and deploy state of the art time-series foundation models on clusters with thousands of GPUs.

• Generate and curate large-scale time-series datasets, real and synthetic, and build evaluators and benchmarks for forecasting and other time-series tasks.

• Push the frontier of Timestral: architectures, training recipes, scaling, and new capabilities such as anomaly detection, classification and embeddings.

• Fine-tune and adapt Timestral on client data, in the cloud or on-premise, and benchmark it against the client's existing models.

• Develop the necessary tools and frameworks to facilitate data generation, model training, evaluation and deployment.

• Collaborate with cross-functional teams to tackle complex use cases, including agents that combine Timestral with Mistral LLMs.

• Manage research projects and communications with client research teams.

• Fluent in English with excellent communication skills, at ease explaining complex technical concepts to both technical and non-technical audiences.

• Expert with PyTorch or JAX.

• Solid knowledge of modern deep learning and large language models: architectures, pre-training, post-training and evaluation.

• Experience pre-training or fine-tuning time-series foundation models or large transformer models at scale.

• Strong background in time-series modelling: probabilistic forecasting, multivariate series and covariates, and classical baselines (statistical models, gradient boosting).

• Not afraid of contributing to a big codebase; finds their way around independently with little guidance.

• Writes clean, readable, high-performance, fault-tolerant Python code.

• Doesn't need roadmaps: moves independently and is comfortable with uncertainty.

• Low-ego, collaborative and eager to learn.

• Track record of success through personal projects, professional projects or academia.

Nice to have

• Ex-founder or builder

• PhD / Master in a relevant field (Machine Learning, Statistics, Applied Mathematics, Physics). Exceptional candidates from other backgrounds welcome.

• Research on time-series foundation models: pre-training, scaling, synthetic data or evaluation.

• Hands-on experience training or fine-tuning LLMs, building agents and tool use, or models that combine text and time series.

• Industry experience in energy, industrial operations, telecom or finance, or experience deploying models for industrial clients.

• Experience working on large code bases or large machine learning models at scale.

• Contributions to a large open-source or industry codebase.

• Publications in top academic journals or conferences.

• Enjoys improving existing code: typing, tests, CI pipelines.