Staff Data Scientist (Forecasting)

Source.ag; Amsterdam; Hybride; EUR 120 000 - 150 000; Plein temps

Avantages offerts par ce poste

Hybrid work

Meals provided on office days

Pension contribution

Well-being support

Equipment: 16" MacBook Pro

Commuting travel allowance

Learning budget: €1,000 annually

Home setup budget: €550

Connectivity: €50/month for WiFi/phone

Flexible holiday policy and minimum 25

Quarterly company events

Résumé du poste

Source.ag in Amsterdam is seeking a Staff Data Scientist to own harvest forecast accuracy and to build hybrid models that combine crop physiology with climate data. You’ll work with data scientists, crop scientists and growers to deliver production-ready forecasts.

You will design robust evaluation backbones, conduct backtests, calibrate uncertainty, and ensure models perform reliably in production while mentoring peers and advancing the team’s technical standards.

Qualifications

• Around 10 years of hands‑on modeling or forecasting in real systems.
• Forecasts that were not trusted and turned into business decisions.
• Models that broke with changes in the environment and were fixed.
• Deep expertise in time series and probabilistic forecasting.
• Validation tying error analysis to business impact and drift monitoring.
• Production‑grade Python and solid engineering habits (Git, testing, CI/CD, Databricks, AWS).
• MSc or PhD and first-principles reasoning.
• AI‑native work with AI agents and LLM tools.
• Collaboration with teams and clear communication with stakeholders.

Responsabilités

• Own the accuracy of harvest forecasts for tomato, pepper and cucumber, especially during unusual weeks.
• Decide which problem to fix first, including input data, plant sampling, or growers’ planning.
• Design and build hybrid models that combine crop physiology and greenhouse climate with statistical methods.
• Build evaluation backbone: backtests on past seasons, strong baselines, uncertainty calibration, production monitoring.
• Ship models to production with engineers and maintain their behavior in production.
• Raise the bar of the data science team through review, challenge, and mentoring.

Connaissances

Time series forecasting

Probabilistic forecasting

Bayesian models

Hierarchical models

Gradient boosting

Deep learning for sequences

Uncertainty quantification

Production-grade Python

Git

Testing

CI/CD

Databricks

AWS

Formation

MSc or PhD in a quantitative field

Outils

Python (pandas/polars)

PyTorch or JAX

PyMC or similar

Scikit-learn

Git

CI/CD tooling

Databricks

AWS

Description du poste

The problem

Every week, a tomato grower tells their buyers how many kilos they’ll deliver three weeks from now. Contracts, trucks, and prices are set on that number. Source builds AI software that greenhouse growers in Europe and North America use for decisions like this, and our Harvest Forecast for tomato, pepper, and cucumber is one of the inputs to that weekly number.

Getting an average week right isn’t the hard part. The value is in the weeks when the crop does something unusual: a heat wave, a sharp drop after a peak, the start and end of the season. Those are the weeks when growers need a forecast most, and when forecasts, ours included, are trusted least.

Why this is hard:

• Slow feedback. You can replay past seasons with any model. But a new model in production shows its first real result only three to four weeks later, and the weeks that matter come only a few times a year. You can’t A/B test a crop.
• Thin, noisy data. A few plants, measured by hand once or twice a week, stand in for a whole greenhouse. Sensors drift. Every site, variety, and grower strategy is a little different.
• Not all of it is biology. Part of every swing comes from the plant, and part from decisions the grower makes after the forecast is made. Telling the two apart is part of the job.
• Strong science, messy greenhouses. Decades of crop and climate science exist. Data alone won’t rediscover it, and the science alone doesn’t fit a real greenhouse.
• People act on the number. A forecast that’s confidently wrong costs a grower real money, so the uncertainty has to be honest.

We’re looking for a Staff Data Scientist who has done this before: taken a forecast of a physical system that people didn’t trust, and turned it into a number a business commits money on.

The difficulty isn’t one clever model. It’s the whole system: data quality, how plants are sampled, growers changing their harvest plans, and then the model. There may be ten things to fix before the forecast is good, and only some of them are in the model. You look at the whole system end to end, find the places that matter, do the hardest work yourself, and carry it all the way to something growers use.

This is a senior individual contributor role, technical all the way, with no people management. You report to our Director of Engineering and join our staff engineers, the senior group that works across teams with the CTO and sets the technical bar. Day-to-day, you work with our data scientists, crop scientists, and directly with growers. Harvest forecasting is the first problem. The same kind of problem shows up across our other crop models, and we expect your work to reach those too.

The role is based in Amsterdam, because you can’t see the whole system from a distance. You need to be in the room with the data scientists, the crop scientists and the people who talk to growers every week.

What you’ll do

• Own the accuracy of our harvest forecasts for tomato, pepper and cucumber, especially in the unusual weeks, from diagnosis to a fix growers use
• Decide which problem is worth solving first, including when the fix is in the input data, the plant sampling, or the grower’s plan rather than the model
• Design and build hybrid models that combine crop physiology and greenhouse climate with statistical and machine learning methods
• Build the evaluation backbone: backtests on past seasons, strong baselines, uncertainty calibration, and production monitoring, so the next ten fixes get cheaper
• Ship your models to production with our engineers, and stay responsible for how they behave there
• Raise the bar of the data science team through review, challenge, and mentoring

We care most about what you’ve done. Today you might be called a data scientist, forecasting scientist, ML scientist, quant, or research engineer. The title doesn’t matter to us.

• Around 10 years of hands‑on modeling or forecasting, with models of physical or real‑world systems that ran in production for more than one season. Energy, weather, process industry, plant breeding, pharma, batteries or trading: the field matters less than the fact that it was real
• You’ve taken a forecast people didn’t trust and turned it into one a business commits money on
• You’ve shipped models that broke when the world changed (a new season, a new site, a new customer) and fixed them, more than once. You can tell us what broke each time
• Deep expertise in time series and probabilistic forecasting: Bayesian and hierarchical models, gradient boosting, deep learning for sequences, and uncertainty quantification on small, irregular datasets
• Validation that holds up: backtests against strong baselines, error analysis tied to business impact, drift monitoring, and evaluation tools other people kept using after you built them
• Production‑grade Python (pandas or polars, scikit‑learn, PyTorch or JAX, PyMC or similar) and solid engineering habits (Git, testing, CI/CD, Databricks, AWS), so your models run in production without you
• An MSc or PhD background and the habit of reasoning from first principles
• You work AI‑native: coding agents and LLM tools are part of how you work every day, and you know where they fail
• You’ve worked alongside people with higher standards than yours, and can tell us what they taught you
• Clear communication with growers, product and engineering alike

Nice to have

• A degree in engineering, physics, applied mathematics, statistics or econometrics
• Experience with physics‑informed ML, mechanistic or process‑based models, data assimilation (for example, Kalman filtering) or system identification
• Publications, talks or open‑source work on forecasting or scientific ML

Is this for you?

You’ll probably like it here if you:

• Want a hard problem that’s still open, not a plan to execute
• Would rather find the real cause than tune a metric
• Stand by your view under pressure, and change it when the evidence says so
• Want to meet the people who use your work and hear directly from them when it’s wrong

You probably won’t like it here if you:

• Need a clear spec, an A/B test, or a large, clean dataset to know you’re right
• Judge your work by an offline benchmark score
• Want to lead a team rather than solve the problem yourself
• Want to work only on the model, and leave the data, measurements, growers and production to someone else

What we offer

• Hybrid work: we work together at the Amsterdam office on Mondays and Thursdays
• Meals: lunch provided on office days
• Pension: a 4.5% pension contribution
• Well‑being: mental health support through OpenUp
• Equipment: a 16\" MacBook Pro
• Commuting: a travel allowance for your commute to the office
• Learning: an annual learning budget of €1,000
• Home setup: a €550 budget for your home office
• Connectivity: €50 a month for WiFi and phone
• Time off: a flexible holiday policy, and we encourage you to take at least 25 days a year
• Community: quarterly company events, including dinners and team activities

How we hire

1. Intro call with the hiring manager or our People Team.
2. One story: tell us about a time a hard project changed direction because of something you did. We’ll ask follow‑up questions to understand your part in it.
3. Workshop on a real problem: we send you a short write‑up on how a greenhouse works, how plants are measured, and where our forecast struggles. Then we work through it together, using your own tools, including AI. We’re not looking for you to agree with us. We want to see what you’d do next.
4. Final conversation and references.

Note for candidates outside the Netherlands: this role is based at our Amsterdam office. After a first video call, we’d love to meet you in person for the next interviews over the next few weeks, and we hope to fill the position as soon as possible. Travel for interviews and relocation are at your own cost.

Note on AI: we use an AI note‑taker during interviews, so we can give you our full attention instead of typing. We don’t record video. If you’d rather we didn’t use it, just let us know at the start of the interview.

Note to recruitment agencies: we don’t accept unsolicited outreach about our open roles. Please don’t contact us about this vacancy.

At Source, we value diversity of background and perspective. People assess their own qualifications differently, so please apply even if you do not tick every box. We are looking for the right person, not a perfect CV.