You're building a Metabase dashboard for a craft brewery outside Barcelona — reconciling revenue definitions across three sales channels and delivering a single source of truth the client checks every Monday morning.
The discipline skills: reconciling conflicting metric definitions across data sources (three systems that define "revenue" differently), building your first BI dashboard in Metabase, designing filters and segmentation views, and validating that every panel on the dashboard uses the same governed definition.
The AI-direction lesson: directing AI to configure a platform is different from directing it to write code. Metabase panels use SQL behind the scenes, and AI will generate different SQL for the same metric across different panels — producing a dashboard where two charts show different revenue numbers. Cross-checking enters here: directing a second AI to review your metric definitions with fresh context surfaces gaps the first AI normalized away.
Your Role
You're building the dashboard. Same tools as last time, plus two new ones: Docker and Metabase. Docker runs the dashboard platform. Metabase is where you build the dashboard itself — a BI tool instead of matplotlib code.
Your verification toolkit grows too. Last time you used AI self-review. This time, you also direct a second AI to cross-check your metric definitions — fresh eyes on the most consequential decisions in the project.
What's New
Last time, you authored metric definitions for the first time and cleaned messy data across three sources. You learned what it means to define "revenue" and document every cleaning decision.
This time, three systems define "revenue" differently and you have to reconcile them into one governed definition before anything else can happen. The reconciliation is the foundation. Every chart on the dashboard inherits whatever definition you choose.
The dashboard is a new medium. You've built charts in Jupyter. Metabase is a different paradigm: configuration instead of code. The underlying concerns are the same; the way you work with AI changes.
The hard part is consistency. A dashboard where two panels show different revenue numbers because they use different SQL behind the scenes destroys trust. Carmen won't know which number to believe — and neither will her sales manager.
Tools
- Python 3.11+ (via Miniconda, "analytics" environment)
- DuckDB
- Jupyter Notebook
- pandas
- matplotlib / seaborn (for exploratory analysis)
- Metabase (new — BI dashboard platform, via Docker)
- Docker (new — runs Metabase)
- Claude Code
- Git / GitHub
Materials
- Four data files — taproom POS transactions, distribution invoices, WooCommerce orders, and a product catalog with canonical names and costs per liter. The product names are different in every system.
- Analysis template — sections for data quality assessment, metric definitions, segmentation findings, and cleaning decisions. You fill it in as you work.
- Dashboard layout template — guidance on where KPIs, trends, and breakdowns go on the dashboard.
- Docker Compose file — starts Metabase with one command.
- CLAUDE.md — project governance file with client context, work breakdown, and verification targets.