Automating carbon-footprint scoring for a climate-impact food platform

Approach

We started from the gap between how sustainability analysts actually score a product — ingredients, sourcing, packaging, process — and how quickly a shopper expects to see a colour-coded answer in the aisle. Hand-scoring doesn’t survive contact with a supermarket catalogue, so we framed the problem as a retrieval task: match each product’s metadata against a reference base of lifecycle-impact data, then fall back to model-based inference when the match is weak. That let us keep the sustainability team in the loop on edge cases without making them the bottleneck on everything else.

Solution

We shipped an ingestion pipeline that takes raw product feeds (titles, ingredient lists, category tags, weights, packaging hints) and resolves them to structured impact factors, with confidence scores that flag low-certainty items for human review. A retrieval layer grounds each score in a reference corpus of ingredient- and category-level lifecycle data, so the output is explainable rather than a black-box guess. We wrapped it in an API the GreenSwapp team could offer to supermarkets and online grocers, and a simple web surface for the consumer side — barcode lookup, low/medium/high impact bands, and swap suggestions built on the same scoring backbone.

What changed

Carbon scoring stopped being a bottleneck on GreenSwapp’s roadmap. Instead of analysts grinding through products one by one, the pipeline now handles whole-catalogue ingestion and hands back explainable, retailer-ready scores — which is what made the platform plug-in-able for grocery partners in the first place. The team could focus on what their reference data should cover next, rather than on keeping up with the next catalogue drop.