You're building a recommendation feature pipeline for a sustainable fashion marketplace in Berlin -- versioned features with DVC, embedding-based product representations from sentence-transformers, and a production monitoring dashboard that serves people who didn't build the model.
The discipline skills: versioning data transformations alongside code with DVC so any historical experiment can be reproduced, generating dense vector features from product descriptions using sentence-transformers (learned representations as inputs to a classical scikit-learn pipeline, not an LLM), separating feature computation from model training, and designing monitoring that tracks disaggregated production metrics with ground truth delay awareness.
The AI-direction lesson: you will encounter territory where you don't know what to verify. DVC integration, embedding boundary discipline, and ground truth delay are all new. Meta-prompting -- asking AI to help design the verification plan while you evaluate whether the plan is comprehensive -- becomes the skill. AI generates feature engineering code without DVC integration, computes embeddings on the full corpus before splitting, and builds monitoring dashboards with technically correct charts that are unintelligible to non-technical stakeholders. When you don't know the right answer, you can still direct AI to help you figure out what to check for. The judgment is whether AI's verification plan covers the failure modes that matter.
Your Role
No guides this time. You receive a brief with deliberate gaps, a dataset with embedded complexity, and a DVC configuration template. You synthesize the brief into a project plan, decide what features to build, choose how to represent product descriptions, design the monitoring system, and pick which metrics to track.
What's New
Last time you built CI/CD infrastructure — eval gates and drift detection. The model was already working; the challenge was keeping it reliable.
This time you build the features that the model depends on. DVC enters as a tool for versioning data transformations alongside code — so you can trace which features produced which results at any point in the project's history. Sentence-transformers enter as a way to convert product descriptions into dense vector features that feed a classical scikit-learn classifier. This is not using an LLM. It is using learned representations as features in a conventional pipeline.
The monitoring deepens too. Drift detection told you when inputs changed. Now you need to know if the recommendations are actually correct — and that answer may not arrive for weeks. Your monitoring dashboard serves people who did not build the model.
The hard part is not any single technique. It is managing the interactions: features that need versioning, embeddings that need boundary discipline, monitoring that needs to communicate, and a client brief with gaps you have to discover through conversation.
Tools
- Python — scripting, feature engineering, monitoring
- DVC — data and feature versioning (new)
- sentence-transformers — embedding generation (new)
- scikit-learn — modeling, evaluation (familiar)
- pandas — data manipulation (familiar)
- MLflow — experiment tracking (familiar)
- Git / GitHub — version control (familiar)
- Claude Code — AI direction, plan mode (familiar)
Materials
You receive:
- Transaction data covering three years of purchases, browsing, wishlists, and returns
- Product metadata with descriptions, sustainability certifications, and seasonal tags
- Customer profiles with demographics and preferences
- A sample of production recommendation events for monitoring work
- A DVC pipeline configuration template
- A project governance file (CLAUDE.md)