Onlyone, eco-responsible banking, Paris
From receipt to product-level carbon
Alongside the Eco Coach, I designed and built Onlyone's receipt scanning system, users photograph a retail ticket, and the pipeline extracts products, normalizes them against reference databases, and estimates carbon impact at the item level.
Transaction-level carbon estimates are useful, but imprecise. A supermarket payment could be groceries, cleaning products, or both. Receipt OCR closes that gap by reading the actual ticket: every line item becomes a product with a name, a price, and an estimable environmental footprint.
Retail receipts are noisy: truncated product names, inconsistent formatting, thermal paper artifacts, and store-specific abbreviations. Turning a raw OCR output into a reliable carbon estimate required a full data pipeline (cleaning, normalization, matching, and scoring), not just a vision API call.
From photographed receipt to per-product footprint: a GCP pipeline that reads, cleans, matches, and scores every line before surfacing it in the app.
Open carbon and product data matched to every normalized receipt line.
ADEME
General carbon emission factors for goods, services, and transport, the baseline for non-food items.
Open Food Facts
Open food product database with composition and environmental labels, used for grocery receipt lines.
Turn noisy OCR output into usable product records, ready to match against reference catalogs.
Transforms raw OCR text into structured product records, parsing line items, inferring products from short names, and normalizing labels via fuzzy matching and embedding-based similarity to bridge the gap between receipt abbreviations and catalog entries.
Match normalized products against ADEME and Open Food Facts, then break down the basket's footprint.
Matches normalized products against reference databases (ADEME for general carbon factors, Open Food Facts for food items), using similarity algorithms to find the closest credible emission factor for each line.
Statistics and visualizations by supply chain stage (distribution, production, packaging, transport), so users see where the carbon in their basket actually comes from.
Surface confidence levels for each estimate and suggest more responsible alternatives.
A confidence score on every estimate reflecting match quality, data completeness, and source reliability, the app surfaces high-confidence results prominently and flags uncertain ones for user review.
Product-level recommendations suggesting less carbon-intensive or local alternatives for items on the receipt, actionable swaps grounded in the matched reference data.
Receipt OCR complements transaction-level estimates with granular, item-by-item insight, the most accurate carbon picture Onlyone can offer without manual user input.
Every product on the ticket gets its own footprint, not an average inferred from the merchant category.
Users see when an estimate is solid and when the system needs help, building trust through transparency, not false precision.
Recommendations tied to actual purchased products, local, lower-carbon alternatives users can try on their next shop.
End-to-end flow from camera to carbon breakdown in seconds, ML APIs on GCP designed for the mobile app's latency and reliability requirements.
GCP Cloud Vision OCR, Python data pipeline, FastAPI and Flask APIs, BigQuery, Tableau analytics, GitLab CI/CD
GCP
BigQuery
Python
FastAPIFlask
Tableau
GitLab CI/CDReceipt OCR gave Onlyone the missing layer of precision, turning a crumpled ticket into structured, scored, actionable carbon data, one product at a time.