Meet Percolate and the toolkit

You'll build against Percolate, a fictional direct-to-consumer subscription coffee company. The data is fabricated but coherent: it spans 2024–2025 and behaves like a real business — revenue grows over time, paid channels cost more to acquire customers, some subscribers churn.

The starter kit

Everything is in course-assets/semantic-agent/. The data lives in a SQLite database — a single file, no server to run — so you can query it from anywhere, and the agent can too.

data/
  schema.sql          # the raw operational schema
  generate_data.py    # deterministic data generator
  percolate.db        # the built database
semantic_layer/       # you dissect and extend this in Part 1
agent/                # you build this in Part 2
app/                  # you deploy this in Part 3

The schema (six tables)

• customers — one row per customer: signup_date, country, region, acquisition channel, segment (Consumer/Pro).
• products — the catalog: name, category, price, cost.
• orders — order headers: customer, date, platform, status (completed/refunded).
• order_items — order lines (the grain for revenue): product, quantity, unit_price, discount.
• subscriptions — recurring plans: plan, start/cancel dates, mrr_amount, status.
• marketing_spend — monthly ad spend per channel.

Notice the shape: two fact tables you'll aggregate (order_items, subscriptions), dimension tables that describe them (customers, products), and a standalone spend fact. This is a classic analytics layout — and exactly the kind of structure a semantic layer is built to tame.