Grocery Swiper is a three-step pipeline that turns weekly grocery flyers into a swipe-based preference dataset and a personalized recommendation email. The project is split into:
- data collection + processing
- a Tinder-like swipe app for labeling
- a recommender + email notification system
Blog context (Part 1-3):
- https://marcrulo.github.io/MonthlyProjects/grocery-swiper-part1
- https://marcrulo.github.io/MonthlyProjects/grocery-swiper-part2
- https://marcrulo.github.io/MonthlyProjects/grocery-swiper-part3
Goal: build a growing dataset of grocery items from weekly flyers.
What happens:
- Scrape flyer items from tilbudsugen.dk, including hidden metadata exposed in JSON payloads.
- Save raw data as CSV, download images, and create enriched item features.
- Processed outputs are stored alongside raw snapshots for historical tracking.
- A GitHub Actions workflow runs the scrape + processing on a weekly schedule to keep the dataset fresh.
Where in code:
- scripts/scrape.py: scraping logic for weekly offers.
- scripts/processing.py: feature processing and image handling.
- data/csv/raw and data/csv/processed: saved weekly snapshots.
Goal: collect like/pass/super-like labels from real users.
What happens:
- A lightweight web app mimics a mobile swipe experience.
- Users swipe on items to generate preference labels that are stored in a local database.
- The UI keeps history for revisions and includes a full-screen mode to feel more app-like.
Where in code:
- app/app.py: Flask app and API endpoints.
- templates/index.html: app layout.
- app/static/script.js: swipe logic + client-side behavior.
- app/static/style.css: UI styling.
- sql/init_db.py: database initialization.
Goal: learn preferences from the swipe data, rank new items, and send weekly recommendations.
What happens:
- Train a preference model on engineered features (price + categorical fields).
- Use active learning to surface uncertain items for future swipes.
- Rank new, unseen groceries and generate an HTML email with the best picks.
- The recommendation script runs after the weekly scrape as part of the GitHub Actions workflow.
Where in code:
- scripts/recommender.py: ranking, recommendation, and email sending workflow.
- Scrape weekly flyer data -> raw CSV + images
- Process data -> cleaned features + processed CSV
- Swipe app -> preference labels stored in SQLite
- Recommender -> ranked items
- Email -> weekly recommendations