End-to-end mini-applications built on the same techniques the courses teach — risk scoring, quant trading, knowledge graphs, macro pipelines. Each case ships with the dataset, the model, and a one-page write-up of the reasoning behind it.
Seven short cases for the AI-era classroom: students self-study the technical content before class, then use class time to diagnose, challenge, revise, and submit a decision artifact before leaving.
Teams inspect a messy business dataset, define the decision question, classify anomalies as error / risk / opportunity, and decide what data can be trusted before modelling begins.
Submit · decision question · data risks · anomaly actions
Build a first predictive model, test it on held-out data, and decide whether performance reflects signal, leakage, overfitting, or a validation design that does not match the business decision.
Submit · target · validation design · trust verdict
Move from prediction to action. Teams draw a DAG, name treatment, outcome, confounders, and colliders, then decide whether the evidence can support a causal business recommendation.
Submit · DAG · identification concern · decision limit
Compare forecasts under uncertainty, choose a horizon that matches the decision, define downside scenarios, and write a monitoring rule for when the model should be revised or retired.
Submit · forecast · uncertainty range · trigger metric
Turn unsupervised output into managerial action. Teams name segments, design actions, check stability, and defend whether the clustering is useful or just an algorithmic artifact.
Submit · segment names · action plan · stability concern
Use embeddings or LLM-aided extraction to create a feature, then audit prompt versioning, model drift, validation samples, cache keys, and whether the feature is safe enough to enter a production model.
Submit · AI feature · validation sample · use/reject decision
Teams prepare a final recommendation, model-card evidence, calibration or fairness concerns, monitoring plan, and deployment verdict: deploy, revise, or reject.
Submit · model card · risk memo · deploy/revise/reject
Mini-applications built on the same methods taught in the courses, with datasets, models, and concise reasoning notes.
Patient-level risk scoring on tabular EHR data. Gradient-boosted ensemble with calibration, SHAP attributions, and a fairness audit across protected groups.
Stack · pandas · lightgbm · scikit-learn · SHAP
Classic cross-sectional momentum portfolio — formation / holding period sweep, transaction-cost-aware backtest, Newey-West t-stats, and a sub-period regime breakdown.
Stack · pandas · numpy · statsmodels
Pair selection by Engle-Granger cointegration, half-life-aware OU-process entry / exit thresholds, walk-forward live simulation with slippage and stop-loss controls.
Stack · numpy · statsmodels · backtester
Schema → entity-resolution → graph store → retrieval-augmented generation over an organisation's internal docs. Includes a Neo4j + pgvector reference architecture and an evaluation harness.
Stack · NetworkX · rdflib · Neo4j · pgvector
Live macro indicators (BIS, IMF, FRED) pulled via SDMX into a dashboard with consistent vintages, revision-aware time series, and nowcasting on key headline series.
Stack · pandasdmx · duckdb · plotly
End-to-end social-media monitoring: ingest → claim extraction → stance classification → community detection on the share graph, with a small operator UI for triage.
Stack · spaCy · transformers · NetworkX
Short-horizon nodal-price spike forecaster. Combines weather, unit-commitment fundamentals, and a gradient-boosted residual to beat naive seasonal benchmarks on out-of-sample tests.
Stack · pandas · lightgbm · joblib
Live calculator for two-sample tests on proportions and means. Handles unequal allocation, sequential / fixed-horizon designs, and plots power curves you can paste straight into a planning doc.
Stack · scipy.stats · numpy · plotly