The client
Spotem runs a corporate gifting business whose customers often enquire with photos.
The challenge
Enquiries often arrived as photos, screenshots or reference images rather than product names.
Staff compared images against hundreds of catalogue products by hand.
Keyword search could not find visually similar products, and results depended on who was searching.
How a photo becomes a match
- 1Upload
Staff upload the customer’s image securely
- 2Analyse
Amazon Nova Lite analyses the image
- 3Embed
Titan Multimodal Embeddings turn it into a vector
- 4Match
pgvector with HNSW finds the closest products
What we did
- 1Discovery
studied image-based enquiries; chose Amazon Bedrock with Amazon Nova Lite and Titan Multimodal Embeddings so there was no computer-vision infrastructure to manage.
- 2Design
serverless architecture with API Gateway, Lambda, S3, Aurora PostgreSQL Serverless v2 and Bedrock; vector search with pgvector and HNSW indexing.
- 3Build
secure image upload, AI image analysis, multimodal embeddings, and a recommendation service.
- 4Security
IAM, Secrets Manager, RDS Proxy, private networking and VPC endpoints; everything provisioned with CloudFormation.
- 5Optimisation
faster search with HNSW, Aurora tuned for workload-based scaling, lower networking cost with VPC endpoints.
The architecture
- Amazon Bedrock
- Nova Lite
- Titan Multimodal Embeddings
- Aurora Serverless v2
- pgvector
- Lambda
The result
Staff upload a customer’s reference image and get visually similar products within seconds.
Recommendations no longer depend on one person’s memory of the catalogue.
Want something like this for your team?
Tell us the questions your team asks or the work that slows you down. An engineer will suggest a small first build on Amazon Bedrock.
