Seeing food waste before it costs money

A Retail business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas.

The pipeline pairs two kinds of AWS AI. Amazon Bedrock provides a large multimodal model that reasons over an image alongside context, the item's weight, the customer's available products, and what has already been logged, to return the most likely matches in plain terms. Amazon Rekognition adds a custom vision model trained on the client's labelled food images for sharper recognition. A hybrid design uses both intelligently: during busy rush-hour periods, an image is first analysed by the custom Rekognition model and its output is fed into the Bedrock prompt for a final decision, while outside rush hours the image goes straight to Bedrock, which keeps quality high when it matters and cost low the rest of the time.

The proof of concept demonstrated an end-to-end, automatic food-recognition pipeline: image in, ranked identification out, every event logged for audit, with the hybrid Rekognition-and-Bedrock design and automated model scheduling proving that accuracy and cost control could go together. It gives the client the evidence and the architecture to take the capability toward production. Figures below reflect the design assumptions set out in the scope of work.

• Recognition that works two ways — A custom Amazon Rekognition model and Amazon Bedrock combine during peak periods, with Bedrock alone outside them and an automatic fallback to Bedrock if the Rekognition model is unavailable.
• No double-counting — DynamoDB session grouping by customer and day, with predictions excluding already-confirmed items, keeps the waste tally accurate and resets cleanly each day.
• Automatic, end to end — An upload to API Gateway triggers the whole flow through Lambda, with no manual intervention, and every image and inference recorded for audit.
• Cost controlled by design — EventBridge and Systems Manager run the custom vision model only during rush hours, and serverless, usage-based services keep the projected running cost low.
• Handover — Cloud Combinator delivered the pipeline as CloudFormation with demonstrations, documentation, and Loom walkthroughs, plus guidance on the steps to take the service to production.

With feasibility proven on AWS, the client has a clear, cost-aware foundation to build on, extending the recognition pipeline toward a production service that helps its food and beverage customers cut waste and cost.