Senior Data Scientist - (Content, Consumer)

Delivery Hero; Berlin; Hybrid; EUR 120.000 - 180.000; Vollzeit

Benefits dieser Stelle

Hybrid work model

27 days holiday

Educational budget

Language courses

Parental support

Udemy Business access

Health checks

Gym access

Employee share purchase plan

Sabbatical bank

Public transport discount

Life & accident insurance

Corporate pension plan

Meal vouchers

Zusammenfassung

Delivery Hero seeks a Senior Data Scientist to own the LLM systems behind social proof across multilingual platforms. You will shape architecture, model strategy, and evaluation, setting the LLM bar for a squad that handles high-scale content and user-generated data.

You will drive problem discovery, quantify business impact, and translate insights into scoped initiatives, balancing cost of inference with quality. Hybrid work supports collaboration at our Berlin campus two days weekly.

Qualifikationen

• Produced LLM-based systems in production with multilingual Text data.
• Ability to reason about cost, latency and quality across models.
• Experience with offline evaluation and LLM-as-judge patterns.
• Observability, tracing and cost tracking in complex pipelines.
• Turn ambiguous needs into scoped DS problems and MVPs.

Aufgaben

• Own the end-to-end LLM systems behind social proof across languages and platforms.
• Define model/provider selection and systematic prompt strategy.
• Drive roadmap via problem discovery with quantified business value.
• Lead productionization and feedback loops with backend/data engineering.
• Elevate team practices through mentoring and ownership culture.

Kenntnisse

LLM systems depth

Multilingual data

Model strategy

Evaluation rigour

Observability

Product thinking

MVP mindset

Mentoring

Tools

Python

SQL

CI/CD

Monitoring

Data pipelines

Jobbeschreibung

As the world’s pioneering local delivery platform, our mission is to deliver an amazing experience, fast, easy, and to your door. We operate in around 65 countries worldwide powered by tech, designed by people. As one of Europe’s largest tech platforms, headquartered in Berlin, Germany. Delivery Hero has been listed on the Frankfurt Stock Exchange since 2017 and is part of the MDAX stock market index. We enable creative minds to deliver solutions that create impact within our ecosystem. We move fast, take action and adapt. No matter where you're from or what you believe in, we build, we deliver, we lead. We are Delivery Hero.

Job Description

We are on the lookout for a Senior Data Scientist to join our Content tribe.

We are building the ratings and reviews systems — social proof — that shape how millions of people decide what to order, across dozens of markets and languages. It is a greenfield space: the current system is early, and the interesting decisions have not been made yet.

The raw material is the hard kind. Millions of short, noisy, multilingual, contradictory pieces of user-generated text, which have to become something a person can act on in two seconds on a restaurant page. Getting that right is an LLM systems problem: aspect extraction, sentiment, summarisation, model selection across providers, systematic prompt optimisation, and the evaluation infrastructure that tells you whether any change made things better.

We are honest about the situation. The team is rebuilding ownership of these systems during a transition, and a lot is undefined. That is the offer: you will not inherit a technical direction; you will set it — and you will set the LLM bar for a team with the appetite and the room to clear it.

Why is this one different?

• You define the direction. Architecture, model strategy, evaluation approach — these are open questions, and they become yours.
• You are the LLM authority for the squad, not a contributor to someone else's. Part of the job is raising everyone else's ceiling.
• Real scale, real consequence. Your models move conversion across Delivery Hero's global platforms.
• Genuinely unsolved problems. Multilingual UGC at scale, empirical multi-provider model selection, prompt optimisation, and evaluation infrastructure for generative output — none of these have a settled answer here or anywhere.
• Agentic development is a first‑class part of how we work, used pragmatically to move faster without losing system understanding.

Your Mission

• You’ll own the LLM systems behind social proof end‑to‑end — quality, reliability, and coverage across languages, platforms, and use cases — including the unglamorous parts: drift, miscalibration, silent quality degradation, and data issues in production.
• You’ll set the standard for how this team builds with LLMs. Model and provider selection based on empirical cost, latency, and quality evidence. Prompt strategy that is systematic rather than folkloric. The judgment on when an LLM is the wrong tool.
• You’ll drive the roadmap through problem discovery, finding high‑impact gaps, quantifying the business value, and turning them into scoped initiatives — treating cost of inference as a product decision, not only an engineering constraint.
• You’ll define and operationalise meaningful metrics, and build the evaluation infrastructure behind them — offline eval suites, LLM‑as‑judge frameworks, and annotation processes — so that evaluation reflects true business value rather than misleading proxies, and a prompt change or model swap becomes a one‑day decision.
• You’ll take prototypes to production, shaping architecture and data flows with backend and data engineering, and building the feedback loops that let the system keep improving without constant manual intervention.
• You’ll raise the bar beyond your own work, through best practices, mentoring, and a culture of ownership and pragmatism.

Qualifications

• NLP & LLM systems depth. You have built LLM-based systems that reached production, preferably on noisy, multilingual, user-generated text. You can explain the architecture, the failure modes you hit, and why you made the calls you made. You have worked across more than one model family and more than one architecture, and you can reason empirically about cost, latency, and quality rather than defaulting to a familiar vendor or pattern. You know when to prompt, when to add context or retrieval, when to fine‑tune, when a small task‑specific model wins, and when not to use an LLM at all.
• Evaluation rigour for generative systems. You know an online A/B test is too slow and too blunt to iterate an LLM system on. You have designed offline evaluation suites, used LLM‑as‑judge patterns while being honest about their biases, structured annotation for generative outputs, and measured faithfulness, hallucination, and quality at scale.
• Observability for LLM pipelines. You instrument what you build: tracing chained calls, watching token usage and cost, and capturing intermediate outputs so a multi‑step pipeline can actually be debugged. You have a view on where this is non‑negotiable and where it is overhead.
• Product thinking and problem framing. You turn an ambiguous business need into a well‑scoped DS problem, spot the high‑impact opportunity nobody put on the roadmap, and connect your technical work to outcomes people outside data science care about.
• Ambiguity and an MVP instinct. You move from zero to one on imperfect information, ship in increments, and resist over‑engineering.
• Leverage beyond yourself. You improve how the people around you work — practices, standards, rigour — and you are the reason a team gets better at something, not just faster at it.

Also valuable

• Production‑grade Python and analytical SQL, with familiarity with ML lifecycle practices, orchestration, monitoring, and modern engineering standards such as version control and CI/CD.
• A strong foundation in statistics and causal inference, and the instinct to challenge a result that looks too good.
• Experience with data collection and labelling pipelines, including working with annotation teams.
• Agentic development tools used pragmatically — including applying them to model improvement itself, such as automated prompt optimisation or LLM‑assisted annotation — with the judgment to know when to verify the output.

Additional Information

• Make the most of our hybrid working model and join the team for face‑to‑face connection and collaboration in our beautiful Berlin campus 2 days a week
• We offer 27 days holiday with an extra day on 2nd and 3rd year of service
• We will support you in developing yourself and your career growth opportunities: 1.000 € Educational Budget, Language Courses, Parental Support and access to the Udemy Business platform to explore a variety of online courses.
• Get moving and release those wonderful, mind‑boosting endorphins: Health Checkups, Meditation & Gym.
• Cash. Dough. Cheddar. Whatever you call it, we’ll help you with it: Employee Share Purchase Plan, Sabbatical Bank, Public Transportation Ticket Discount, Life & Accident Insurance, Corporate Pension Plan
• The power of getting together over some food is unrivaled. Here are a few ways to help you do that. All the yum: Digital Meal Vouchers and Food Vouchers.
• Wondering what relocating to Berlin is like? In this article, we’ve put together 10 things you should know about moving to Berlin and how Delivery Hero can support you. You can also visit our relocation hub and check out more information about moving to Berlin.

We believe diversity and inclusion are key to creating not only an exciting product, but also an amazing customer and employee experience. Fostering this starts with hiring - therefore we do not discriminate on the basis of racial identities, religious beliefs, color, national origin, gender identities or expressions, sexual orientations, age, marital or disability statuses, or any other aspect that makes you, you.

We encourage you to let us know if you need any accommodations or specific accessibility support to ensure a smooth interview experience—just let us know with an email to our Inclusion Officer at [e-mail].

Severely disabled applicants with equal qualifications will be given preferential consideration.

You're welcome to share your pronouns (he/she/they) right from the start so we can address you respectfully from our first contact.