<aside> 📊

Freshflow — Senior Data Scientist (m/f/d) | Remote (CET ±1) · Berlin / Paris

</aside>

About Freshflow

Today, more than 40% of all food produced globally is wasted, causing $1.2 trillion in losses every year and carbon emissions six times higher than the entire aviation industry. Freshflow is building the AI operating system for the fresh food supply chain, starting with ordering and inventory optimisation for fresh departments.

We already work with some of the largest grocery retailers in Germany and France (EDEKA, REWE, Carrefour, Intermarché, E.Leclerc), meaningfully reducing food waste while improving on-shelf availability. 93% of our order proposals are accepted by store staff. We are growing revenue 400% year over year and just closed our $10M Series A.

You would join a small, senior, dedicated Data Science team, and bring your best work to the hardest forecasting problem in retail and develop cutting edge technology.

The hard and interesting problem you will work on

Fresh is the hardest and most interesting forecasting problem in retail for several reasons. Products decay on the shelf with uncertain, item-specific shelf lives. The same onion arrives under different item IDs depending on supplier and week, and when a customer buys an apple, the transaction often doesn't record the variety. Turning that into a coherent demand signal is a challenge in itself, before any forecasting begins. On top of that, inventory records are rarely accurate: shrinkage, unrecorded waste, and scanning errors mean the system often doesn't know what's actually on the shelf. Every order is a daily tradeoff between waste and availability, and getting it wrong in either direction has immediate, visible impact in the store.

What you'll do

• Develop and improve our distributional forecasting models, working across model architecture, calibration, and coverage. This includes deepening existing production models, consistently monitoring forecast quality metrics, and acting swiftly when performance degrades as well as identifying where current approaches break and what comes next.
• Make meaningful contributions to our inventory simulation and ordering policy framework, actively driving progress on both and seeing improvements through to production.
• Shape technical direction in your focus area, contributing to methodology decisions and holding a high bar for code quality, testability and production readiness across the team.
• Contribute to end-to-end pipeline quality, from input data integrity through forecast output to live order recommendation performance. This means ongoing monitoring, regular backtesting evaluations to assess model impact, and a shared commitment to catching issues before they reach the store.
• Ship production-quality work, writing well-tested, readable, thoroughly reviewed code. Stay at the forefront of agentic AI tooling (Claude Code, Cursor or similar) and use it to work efficiently without compromising on craft.
• Collaborate closely with Customer Success and Engineering to translate store-level operational findings into product improvements.

Qualifications

You have built systems that make decisions under uncertainty with real operational consequences, not just research prototypes. Your background is in at least one of probabilistic forecasting, stochastic inventory simulation, or operations research, with solid working knowledge across the others. Right now the biggest opportunities are in inventory simulation and ordering policy design, so depth there will let you make your mark fastest. If your strength lies elsewhere in this space, we still want to hear from you.

Education: MSc or PhD in a quantitative field. Statistics, Mathematics, Physics, Operations Research, Computer Science, or similar. What matters most is depth of technical thinking, not the specific discipline.

Technical skills:

• Strong Python and SQL on large, messy, real-world data. Production-grade code
• Deep familiarity with distributional and probabilistic forecasting methods (e.g. quantile regression, LGBM with distributional output, GAMLSS-type models, conformal prediction)