ML Engineer - Demand Forecasting

4+ years of experience

Europe

Fulltime, remote

Our client builds an AI inventory planner that brings forecasting, optimization and execution into one product. Built for use across industries, it powers replenishment and production planning for customers in wholesale, retail, manufacturing and food service.

At its core is a proprietary time series foundation model, pre-trained on millions of demand series and billions of data points across industries.

Founded in 2026, our client is a spin-off from a university research group in Germany, where the foundation model originated as a research project. The company is based in Germany and is expanding its engineering team to further develop the model and product.

About the role

Our client is looking for a Machine Learning Engineer. You won’t just fine-tune existing models or build API wrappers: you’ll help them build their own time series foundation model — from designing the architecture and training it from scratch to getting it into production.

As one of their first engineering hires, you’ll work directly with the founders, make core technical decisions and help build the team along the way.

Your work spans the full model development cycle: understanding forecast errors in production, developing and testing new approaches, improving pre-training and adapting the model to specific industries. You’ll take projects from initial experiments through deployment, with the scope to make consequential technical decisions and see how your work improves customers’ planning.

What you’ll work on

Infrastructure: Build efficient data pipelines for large-scale pre-training and inference, from resource-efficient data loading and validation to reliable distributed processing of millions of demand series throughout week-long training runs.
Model architecture: Improve our time series transformer’s efficiency and capabilities. Current areas of interest include sparse covariate handling, text inputs and in-context learning across related products.
Pre- and mid-training: Refine the training setup—scaling approaches, loss functions, optimizers—for training stability and forecast quality. Shape data mixtures and curricula for balanced learning across demand patterns, industries and time scales, including rare patterns and edge cases.
Fine-tuning and post-processing: Develop adaptation strategies for specific industries, customers and applications without losing generalization, and build calibration, bias correction and hierarchical reconciliation into the final forecasts.

Stack: Python, PyTorch, Polars, PostgreSQL, MLflow, Slurm, Docker, Kubernetes, FastAPI.

Requirements for the position:

Fluent in Python, PyTorch and SQL.

You write production-quality code: tested, maintainable, and used by others.
You have taken ML models from prototype to production and kept them running: monitoring, debugging, retraining.
You have trained deep learning models on datasets that don’t fit in memory, on multiple GPUs, and you know how data loading, GPU memory and distributed training affect throughput and reliability.
Upper-intermediate level of English or German.

Nice to have:

Experience in an early-stage startup or building a product or team from the ground up.
Experience modifying transformer architectures or pre-training models, whether in language, vision, time series or another domain.
Experience with fine-tuning, adapters or transfer learning to adapt pre-trained models to new tasks.
Experience with probabilistic modeling, uncertainty estimation or calibration.
Familiarity with time series forecasting or inventory planning is a plus; you learn the domain on the job.

How the team works:

You keep the team informed of progress and challenges, discuss technical decisions openly and ask for help early, sharing what you’ve tried and where you’re stuck. They work through problems together and give each other direct feedback.

What We Offer: