REWE Group — one of Europe's largest retail groups. Production AI and enablement.
REWE Group — production AI delivery, internal GenAI products, and hands-on enablement.
AI Engineer · Cologne, Germany
Nils spends 4,000+ hours with LLMs
0 / 10,000 h
0%
I build production AI systems that make messy operational work measurable, faster, and easier to trust.
I build eval-first LLM systems with structured outputs, tracing, and agentic tooling that stay controllable in production.
Hours gathered during thesis, research, implementation, problem solving, writing code, removing code, and reviewing code.
Estimated from daily routines: ~6h on weekdays, ~2h on weekends, counted from Dec 1, 2023. Not machine-tracked — honest approximation.
Practical AI systems that fit real workflows and stay measurable.
Agentic engineering, evaluation, tracing, and practical adoption.
Peer-reviewed work on context-aware emotion analysis with LLMs.
Research on operationalizing context-aware emotion analysis with LLM-guided pipelines.
About
Timeline.
I build LLM systems that stay boring, measurable, and controllable — because that's where they actually create business value. Research background, production mindset.
Based in Cologne. Currently at REWE Group. Working across production AI, evaluation, and applied LLM systems.
2025 — today
Junior ML Engineer, REWE Group
Production AI applications, agentic engineering, evaluation, and enablement.
2024 — 2025
Working Student, Research & Innovation Engineering
Agentic systems, evaluation, and proof-of-concept architectures.
2022 — 2025
M.Sc. Business Informatics, TU Wien
LLM systems, evaluation, and context-aware emotion analysis.
2018 — 2022
B.Sc. Business Informatics, WWU Münster
Foundation in business systems, software, and applied AI.
Work
Production AI, Evaluation & Enablement.
Switch between Business and Dev in the header to see the same work at the right level of detail.
Cluster 01
Agentic Engineering
Let AI handle the boring parts of software delivery.
- Set up AI assistants for repetitive coding, review, and deployment tasks
- Help teams ship faster and spend less time on mechanical work
- Design patterns that stay reliable in real production environments
- Turn codebases into places where both humans and AI can work effectively
- Agentic harnesses for automated development, deployment, and maintenance cycles
- Claude Code, Codex, OpenCode, and GitHub Copilot as daily tools
- Skills, AGENTS.md, and MCPs as practical building blocks
- Agent-ready repo structures and task patterns for delivery work
Cluster 02
GenAI Applications
Build real products powered by language models.
- Production AI products live across internal workflows and user-facing experiences
- Measure whether an AI feature is actually good enough to ship
- Own the full path from idea to deployed product
- Work with text, images, audio, and combinations of them
- Build AI features that fit real business workflows, not demos
- Shipped multiple production LLM applications across internal tooling and customer-facing flows. ask me for specifics.
- Specialist in LLM evaluation, from gold sets to trace inspection
- Full stack from development to production deployment on GCP
- Structured outputs, embeddings, TTS, vision, and multimodal automation
- Applications where LLMs are core, but still fit real production workflows
Cluster 03
Enablement
Help teams actually adopt AI, not just experiment.
- Advise on where AI creates value and where it does not yet
- Help colleagues use AI tools in daily work with clear guardrails
- Translate hype into concrete, usable operating habits
- Focus on adoption that survives beyond one-off pilots
- AI strategy with a hands-on engineering lens
- Internal enablement on agentic workflows and LLM tooling
- Bridge the gap between AI hype and production reality
- Adoption patterns that hold up after the pilot phase
Tech Stack
Cloud-native stack for building and running LLM applications at production scale, with a strong bias toward evaluation and observability.
Capabilities
- Production LLM application development
- Evaluation pipelines and quality measurement
- Agentic workflows and AI-assisted engineering
- Cloud deployment and data pipelines
- Python
- GCP
- Claude Code, Codex, OpenCode, GitHub Copilot
- LangSmith, Langfuse
- MCPs, Skills, AGENTS.md patterns
- Dagster, dbt, Snowflake
- Agentic Frontend Development
Personal Focus
Understand where LLMs actually create value. Prefer simple, controllable usage over unnecessary complexity. Maintain and improve systems with a mix of human control and agentic harnesses. Make quality evaluable and quantifiable through custom evaluations.
Featured Case Study
Packaging → shop-ready labels.
A production system at REWE Group that turns product packaging photos into structured master data — so the online shop can legally match what customers see in store, and new products reach the shelf faster.
The master data team came to us: manual label transcription wasn't keeping up. Consumer-protection law requires the online shop to match the physical packaging (country of origin, allergens, nutrition). Trend products need to be listed fast. And supplier data is fragmented — there's no universal API for "what is printed on this package". Photos and videos are.
Two tools, shipped by a team of two in six months:
- Extraction pipeline — photos/videos in, structured labels out. Reviewers approve before anything reaches the master data system.
- Goldhamster — a custom evaluation tool. Gold dataset, scoring dashboard, in-place gold editing, re-scoring without re-running the pipeline. Makes quality measurable and the path to higher accuracy concrete.
- ~96% label-level accuracy on real photos from stores and logistics centres — not curated benchmark images
- ask me for unit economics — 25-75x cheaper than market-available solutions.
- Under one minute end-to-end, from upload to extracted labels
- Happy stakeholder, measurable quality, clear path toward full automation
The hard part wasn't the model. The stakeholders own the problem; I own the solution. The real work is the translation layer between them — explaining error modes, accuracy trade-offs, and why "96% with a good review UI" beats "100% someday". Good engineering here means knowing which decisions are mine to make and which ones aren't.
- GCP
- Snowflake
- Gemini (vision)
- Structured Output
- LLM-as-Judge
- Goldhamster (custom eval)
- Agentic-coded UI
Architecture details, evaluation methodology, and lessons from production are best discussed in conversation.
Reach out on LinkedInResearch
Research on Context-Aware Emotion Analysis.
The research thread is the same one I use in production: make LLM behaviour legible, testable, and grounded in something more robust than intuition. That mindset shaped how I learned to build measurable systems.
Publication · ACM CHI EA 2025
Read publication
Context over Categories: Implementing the Theory of Constructed Emotion with LLM-Guided User Analysis
Author
Nils Kluewer
Where
Yokohama, Japan
Publisher
ACM Digital Library
A context-aware LLM pipeline based on Lisa Feldman Barrett's theory of constructed emotion.
Master Thesis · TU Wien
Read thesis
Context over categories: implementing the theory of constructed emotion
Author
Nils Kluewer
Where
TU Wien, Vienna
Publisher
TU Wien Repositum
Full operationalization via the "context sphere" - a user-specific construct for nuanced emotion analysis.
Contact
Let's talk.
Best reached via LinkedIn. Happy to discuss applied AI, agentic engineering, evaluation, or research collaborations.