A 3-4 hour interactive workshop where students combine SQL, machine learning, and causal inference to build and deploy ad-targeting models in a live A/B test competition.

# Install dependencies
pip install -r requirements.txt

# Generate the workshop database
python generate_data.py

# Start the server
python app.py

The server runs on http://localhost:8000. Share this URL (or your machine's IP) with students.

1. CTR is not revenue: "Window shopper" users click everything but never buy. "Researcher" users rarely click but make large purchases when they do. A pure CTR model shows clickbait to window shoppers — high clicks, no revenue.

2. Full-funnel modeling: The winning strategy models P(click) * P(convert|click) * revenue, not just P(click).

3. Clickbait hurts: High clickbait scores boost CTR but reduce conversion rates. This is intentionally baked into the data generating process.

4. Latency matters: Complex models are penalized in the live simulation. If inference exceeds the latency budget, traffic gets randomly assigned instead of using the model's prediction.

1. Open notebooks/student_workbook.ipynb
2. Set SERVER to the instructor's URL
3. Register their team
4. Explore data via SQL (in notebook or at /sql)
5. Train and evaluate models
6. Wrap model in a ScoringModel class with a .predict() method
7. Pickle and upload to the server
8. Watch the live dashboard at /dashboard

generate_data.py    → Creates SQLite database with synthetic e-commerce data
app.py              → FastAPI server (SQL API, model upload, simulation control)
simulator.py        → A/B test engine (scores teams' models against ground-truth DGP)
static/             → Web UI (landing page, SQL explorer, live dashboard)
notebooks/          → Student Jupyter notebook
data/workshop.db    → Generated database (not in git)

Local network: Run on a laptop, students connect via IP address.

Cloud: Deploy to any platform that runs Python (Render, Railway, Fly.io, EC2). The SQLite database is self-contained.

The synthetic data has five hidden user segments with distinct behavioral profiles:

Students don't see segment labels — they must discover these patterns through data exploration.