Notebooks

Executable Jupyter notebooks for HR AI workflows: real code, run against synthetic data, with the actual cell outputs (including charts) saved in the file. Open them directly on GitHub to read the rendered output, or clone the repo and run them yourself. Never run them against real employee data without first reviewing the risk assessment template.

Contents

File	Purpose
hr-qa-agent-demo.ipynb	Working demo of the HR Q&A agent: loads the eval set, defines an opinionated system prompt, walks through 5 representative scenarios (routine, edge-case, adversarial, sensitive, escalation)
attrition-risk-modeling.ipynb	Baseline attrition risk model on synthetic data: logistic regression + gradient boosting, calibration analysis, a three-part fairness audit (four-fifths selection rates, calibration by group, error-rate parity with small-group cautions), HRBP workflow integration
skills-gap-analysis.ipynb	Identify priority L&D investments by mapping employee skill assessments against a competency framework

Running locally

pip install -r ../requirements.txt jupyter ipykernel
jupyter notebook

hr-qa-agent-demo.ipynb runs end to end with no API key: if ANTHROPIC_API_KEY is not set, call_agent() falls back to a hand-authored reference response for each scenario instead of a live model call, so the notebook always produces meaningful output. Set ANTHROPIC_API_KEY and install anthropic to compare live model output against the reference.

Each notebook is self-contained and generates its own synthetic sample data in the first few cells. Replace the sample data generation with your own data source to adapt a notebook for real use, see each notebook’s “Adapting this for your organization” section.

CI executes all three notebooks on every push (see .github/workflows/ci.yml) to catch breakage before it merges.

Not covered yet

Compensation equity analysis (deliberately: see pay equity governance on why that work is attorney-directed) and onboarding effectiveness measurement. Contributions welcome via CONTRIBUTING.md.