Skip to content

Bodega shopping analytics

Turning supermarket invoices into a structured, queryable dataset and a weekly spend digest, described by method, schema, and aggregate outcomes only.

Data & Analytics No personal data datahub-local-workflows

A blank receipt arcing across a dark studio into a receding grid of tiles

Method, schema, and aggregates only; no shopping rows are published.

Problem

Household spending lives in invoice emails and paper receipts spread across supermarket loyalty accounts and photo libraries. Individually they say little. Together they could answer practical questions: which categories cost the most, which products are rising fastest, and how this month compares with the last.

The hard part is not the analysis. It is building a reliable pipeline from unstructured, personal documents without ever publishing the underlying rows.

Approach

The pipeline treats invoices as an ingestion problem and keeps a strict line between raw personal data and published output.

  • Ingestion: invoices are fetched from supermarket email accounts, and receipts photographed into Google Photos are extracted with OCR.
  • Parsing: line items are normalized into a structured schema of store, product, category, quantity, unit price, line total, and purchase date.
  • Storage: parsed rows land in Iceberg tables on the lakehouse, with history retained through the catalog.
  • Transformation: dbt models aggregate spend by category, product, and store over time, and compute period-over-period changes.
  • Delivery: n8n queries the marts through Trino and sends a weekly digest highlighting current spend, the products whose price rose most, and changes in shopping patterns.

Only the aggregation layer and the digest are shared. Raw invoices, receipt images, and line-item rows stay private.

Technologies

  • Apache Airflow for scheduled ingestion and OCR
  • Iceberg on Garage S3 as the storage layer
  • dbt for aggregation models and tests
  • Trino for the query surface used by delivery
  • n8n for the weekly digest and notifications

Outcome

The dataset turns scattered receipts into a consistent spending history and an automatic weekly summary: spend by category for the current month, the products whose prices moved most, and shopping-pattern shifts against previous months. What this page reports is the method and the shape of the output, never the household's rows.

Further reading

Alvaro Santos Andres