You're building a nurse-to-hospital matching model for a staffing company in Bangalore -- scoring placement quality from a mix of structured fields and free-text descriptions, then auditing the model's predictions for fairness across demographic groups.
The discipline skills: constructing a scikit-learn Pipeline with ColumnTransformer for heterogeneous data (TF-IDF on text, preprocessing on tabular), adapting a pretrained language model through transfer learning with frozen base layers, conducting a fairness audit with disaggregated metrics, and choosing feature selection methods that respect the cross-validation boundary.
The AI-direction lesson: context curation becomes a deliberate skill. What you include in each AI session -- the full dataset schema or just the relevant columns, the complete evaluation plan or just the current metric -- shapes what AI produces. AI generates ColumnTransformers with misassigned columns when context is cluttered. It fine-tunes all layers with a uniform learning rate when it should freeze the base. It performs feature selection on the full dataset before cross-validation, creating a subtle form of leakage that Pipelines should prevent. And it computes overall accuracy without disaggregating by subgroup, hiding the fact that a model with 92% overall accuracy can fail 40% of the time for a specific group. The directing act is curating what AI sees so it produces what you need.
Your Role
You design what goes inside the Pipeline -- which columns get which treatment, which fairness intervention to apply, and how to communicate the trade-offs. Templates provide structure for the Pipeline construction and the fairness audit. You fill them with judgment.
What's New
Last time you took a working model and made it deliverable — Docker, scripted pipelines, prediction logging. The model itself was the starting point, not the challenge.
This time the model is the challenge. The placement data mixes structured columns with free-text descriptions, and they need different preprocessing. scikit-learn Pipelines make the wrong execution order impossible by design — a shift from catching leakage by hand (P4) to preventing it structurally. Transfer learning brings a pretrained model's existing knowledge to a task where training data alone is not enough. And the fairness audit surfaces something aggregate metrics hide: a model that works well overall can fail systematically for one group.
Tools
- Python / pandas — data processing, Pipeline construction
- scikit-learn — Pipelines, ColumnTransformer, feature selection, baseline models
- Hugging Face Transformers — transfer learning (new)
- PyTorch — training loop (familiar from P4)
- MLflow — experiment tracking (familiar)
- Jupyter — development
- Claude Code — AI direction
- Git / GitHub — version control
Materials
You receive:
- The placement dataset — 2400 records with structured fields and free-text nurse bios and hospital notes
- A Pipeline template showing ColumnTransformer structure with placeholder column groups
- A fairness audit guide covering disaggregated evaluation, fairness metrics, and intervention options
- A project governance file (CLAUDE.md)