Executive Summary
Wine-Searcher is the world’s most visited wine marketplace and price comparison platform, connecting consumers, collectors and trade buyers with retailers in more than 180 countries. With over 8 million distinct listings, it depends on accurate, richly detailed product data to power search, recommendations and pricing intelligence.
Peritos Solutions developed an AI-powered label recognition and provenance platform for Wine-Searcher using Amazon SageMaker and Amazon Bedrock. It extracts winery details, vintage, grape variety and certifications from label images and generates historical provenance summaries. Manual work that took up to 20 minutes per SKU now takes under two seconds, across more than 60,000 new labels a month.
Project outcomes
Results & Impact
< 2 sec
Inference latency per label
94%
Label identification accuracy
8M+
Wine listings enriched
60%
Faster data ingestion vs manual
Background
Wine-Searcher’s editorial team relied on contributor-submitted data and manual verification. This created bottlenecks as the catalogue grew and left most listings without historical or regional context. Leadership set a clear brief for AI automation: recognise the label, extract the facts and tell the story.
Wine-Searcher already ran a modern data pipeline on AWS. Peritos was engaged to add an end-to-end AI layer on top, covering image ingestion, a custom recognition model, OCR extraction, RAG-grounded summaries and integration back into live listings. The targets were accuracy above 92% and latency under two seconds per label.

Challenge
Labels vary widely in typography, and regulatory text overlaps with brand content. Many marketplace images were low-resolution, tilted or partially obscured, and labels appeared in six languages.
Older bottles often carried multiple dates, making the vintage ambiguous. Small producers had few training images, and early Bedrock outputs included plausible but fabricated château histories.
On top of this, more than 8 million existing listings needed to be enriched retroactively without excessive cost.
Solution
A custom CNN model, trained on more than 500,000 labelled images and hosted on a versioned SageMaker endpoint, identifies key fields and assigns each a confidence score. Amazon Textract then extracts all label text, and a Node.js Lambda function parses it into Wine-Searcher’s listing schema.
The structured data drives a RAG pipeline. Relevant passages from a curated wine knowledge base, indexed in OpenSearch, are passed to Amazon Bedrock (Claude) to generate grounded provenance summaries of 150 to 250 words.
Key Solution Capabilities
- Label Recognition: Identify winery, wine name, vintage, appellation, grape variety and alcohol content.
- Structured Extraction: Convert label text into validated JSON for live listings.
- Multi-Language Support: Detect and translate French, Italian, German, Spanish and Portuguese labels.
- Grounded Summaries: Generate provenance stories traceable to indexed knowledge base sources.
- Human Review Queue: Route low-confidence outputs to editorial specialists before publishing.
- Continuous Learning: Retrain models automatically from validated corrections.
- Drift Monitoring: Detect accuracy degradation with SageMaker Model Monitor.
Approach
Peritos delivered the platform in six phases: discovery, data pipeline, SageMaker model, RAG and Bedrock, API and listings integration, and UAT and go-live with hypercare.
Augmentation pipelines made the model robust to poor image quality, and a rule engine resolved vintage dates using regional labelling conventions. Hallucination guardrails strip any claim not traceable to an indexed source. A SageMaker batch transform job on spot instances enriched 8 million legacy listings in 72 hours, with checkpointing to handle interruptions.
All inference runs within Wine-Searcher’s AWS account, secured by Cognito, WAF and end-to-end encryption. Enriched listings show 34% higher click-through and 22% higher add-to-cart rates, contributing to an estimated USD 10 million or more in annual value.









