Data Scientist

Adsquare GmbH; Berlin; Hybrid; EUR 92,000 - 107,000; Full time

Benefits offered by this job

Hybrid work

Remote 3 months/year

Learning budget 1200 EUR/yr

30 vacation days

Urban Sports Club

Company pension scheme

Relocation / home office budget

Latest hardware and tools

Job summary

Adsquare GmbH ищет высококвалифицированного дата-сайентиста для разработки и внедрения моделей измерения причинного эффекта в рекламных кампаниях. Вы будете работать в команде Data Science над сложными задачами с геопространственными данными и аудиторией, внедрять модели для оценки влияния кампаний на визиты и поведенческие метрики.

Требуется опыт в Bayesian подходах, временных рядах и обучении LLM. Предлагается гибридная занятость, обучение и современные инструменты в AWS ecosystem.

Qualifications

• 7+ лет hands-on data science с моделями в продакшене и бизнес-эффектом

Responsibilities

• Вы будете научной совестью наших продуктов: перевод миллиардов локатор сигналы в рабочие модели

Skills

Data science опыт

Causal inference

Production ML

SQL / Python

Geospatial data

LLM инструментари

Bayesian statistics

Time series forecasting

Neural networks

Communication

Business acumen

Values: Drive, Resilience, No BS

Education

Магистр/PhD в математике, статистике или естественных науках

Tools

Python

SQL

AWS

Snowflake

dbt

Docker

Terraform

PyMC

Stan

NumPyro

PyTorch

Job description

Описание

Intro Adsquare does Location Intelligence for programmatic advertisers. We solve hard problems with good people, good technology and a lot of data. Our core values guide how we work, and we hire for them:

• Drive: we turn ambition into action that produces valuable outcomes.
• Resilience: we adapt, persevere, and grow stronger through challenges.
• No BS: we are honest, transparent, and clear.
• Humble: we choose modesty over vanity and let results speak.
• Moral Compass: we act with fairness, integrity and respect.

Data and evidence drive every decision we make. A claim that has not been tested on data it has never seen is a hypothesis, not a result. In practice, that means:

• Business need first. Every algorithm starts with agreed-upon business-relevant success metrics before work begins.
• Pre-registration. Hypotheses, primary metrics, analysis plan, and stopping rules are written down before the data is looked at.
• Statistical power. Sample sizes and minimum detectable effects are calculated up front. We do not run experiments that cannot answer their question. And we aim to build ML models with appropriate sensitivity, specificity, and precision.
• Independent tests, honest uncertainty. Every claim is tested on held-out data never used for model selection, and reported with intervals and its assumptions stated.
• Reproducibility. Code, data versions, and seeds are recorded, so anyone, human or agent, can re-run the result.
• Negative results count. A well-run experiment that kills an idea is a success.

You will own this standard, teach it, and raise it.

Задачи

You will be the scientific conscience of our data products: the person who turns billions of location signals into models that work in production and can prove that they work. You own the data science of a whole domain, from the business question to the live model and its measured impact. You set the methodological standard for the team, review the work of other data scientists, and decide with Product which questions are worth answering and formulate answers always with the necessary scientific evidence. You will work in the Data Science team, which is part of the Data Department, alongside other data scientists and ML engineers. Using geospatial data and audience attribute data, your tasks include developing and evaluating algorithms for causal impact measurement. We ask questions like: how much did a campaign change real-world visits, which audiences does a place attract, what will footfall look like next month, and how can AI help to answer these questions faster and better. Build and ship models that answer business questions. Design, validate, and deploy models for causal campaign measurement, audience classification, visitation forecasting, place and audience embeddings, and recommendation. Every model starts from a business need and ends as a monitored production service. Measure causal impact. Own methods that estimate what advertising actually changed: counterfactual and synthetic-control models, Bayesian structural time series, difference-in-differences, and uplift models. Quantify the uncertainty and state the assumptions. Forecast. Build time series models for footfall, reach and audience size, from hierarchical Bayesian models to neural forecasters, and backtest them honestly. Work with representations. Learn vector embeddings of places, devices and audiences; use them for similarity search, collaborative filtering and lookalike modelling. Put LLMs to work on our data. Fine-tune and evaluate LLMs for our own use cases (classifying places and audiences, extracting structure from text, agentic data workflows), with proper evaluation sets, not anecdotes. Run science with agents. Use coding agents such as Claude Code or Cortex to write analysis code, run experiment grids and draft reports, and design the guardrails that keep agent-run experiments reproducible and correct. Set the standard. Write the team's methodological playbook: experiment design, power analysis, validation protocols, review checklists. Review other data scientists' work rigorously and fairly, and mentor senior and mid-level colleagues. Shape the roadmap. Think 12 to 24 months ahead. Bring state-of-the-art research into our products, identify high-value R&D opportunities, and lead solutions that span several projects. Explain it. Write clear technical papers and internal reports, and explain complex algorithms to Product, Sales, and leadership in terms of decisions, risks, and money. Represent Adsquare in peer-reviewed publications and at conferences.

Требования

• Experience: 7+ years of hands‑on data science, or 4+ years after a PhD, with several models you designed running in production and a measured business effect you can name.
• Mathematical depth: you know the maths behind the methods you use: probability theory, inferential statistics, linear algebra, optimisation and experimental design. You can derive an estimator, spot a violated assumption, and explain why a model fails.
• Scientific rigour: you work by the scientific method. You have designed pre‑registered experiments, done power analyses, built leakage‑free validation schemes, and changed your mind when the data said so.
• Causal inference: proven experience estimating causal impact from observational and experimental data, for example synthetic control, Bayesian structural time series, Geo‑Lift, difference‑in‑differences, or uplift modelling.
• Core ML toolkit: production experience with classifiers, vector embeddings and similarity search, and collaborative filtering or other recommender methods.
• Agentic coding (must‑have): you work daily with coding agents such as Claude Code or Cortex, use them to run data science experiments, and know how to check and review their output.
• Large, messy data: you have worked with TB‑scale, complex data in SQL and Python (for example, Athena/Trino, Spark, BigQuery, or Snowflake) and know how to make it fast and correct.
• Production mindset: Python at a professional level (typed, tested, reviewed), plus experience deploying and monitoring models in a cloud environment, ideally AWS.
• Communication: excellent scientific writing, and the ability to explain a complex algorithm to a business stakeholder so they can decide on it.
• Business acumen: you link model metrics to revenue, cost, and customer outcomes, and you know when a simpler model is the better business choice.
• Values: you live Drive, Resilience, No BS, Humble, and Moral Compass, and you care about privacy when working with location data. None of these is required on its own, but the more you bring, the better the fit.
• Geo‑spatial data science (very desirable): hands‑on work with H3 or S2 grids, spatial joins, coordinate reference systems and mobility data, and predictive models on top of them, such as footfall attribution, visitation prediction or place clustering.
• Bayesian statistics: hierarchical models, prior and posterior predictive checks, and tools such as PyMC, Stan or NumPyro.
• Time series forecasting: state‑space and structural models, hierarchical forecasting, and neural forecasters, with honest backtesting.
• Neural networks: you design and train deep models in PyTorch / TensorFlow / Keras and understand why they work, including representation learning.
• LLM‑based data science: you have built LLM solutions on real data and fine‑tuned models (for example, LoRA or other parameter‑efficient methods) for a specific use case, with an evaluation set that proves the gain.
• Education: a Master's degree, ideally a PhD, in mathematics, physics, statistics, or another natural science or quantitative field.
• Publications: first‑author or significant contributions to peer‑reviewed scientific publications.
• Domain: experience in programmatic advertising, ad measurement, or location data.

Our Software Stack Languages: Python, SQL Data and cloud: AWS (Athena, S3 with Iceberg tables, SageMaker, Bedrock, EKS), Snowflake, dbt, Docker, Terraform Science and ML: scikit‑learn, PyMC, FastAPI, Streamlit, Keras, etc. AI tooling: Claude Code and other coding agents, LLM APIs, MCP servers

Условия

On top of a competitive package: Hybrid work, and up to 3 months a year working remotely from anywhere A personal learning budget of 1,200 EUR a year 30 vacation days a year Urban Sports Club membership, company pension scheme and relocation / home office budget Regular team and company events organised by our People team Latest hardware and all tools you need to thrive Salary Range (Annual On‑Target Earnings) Annual On‑Target Earnings (base + bonus) - 92,000 - 107,000

Work model

Hybrid Desired start date January 2027