Portfolio case study · Financial services · Claude · RAG
A recommendation platform scaled from prototype to production, using Claude and Qdrant retrieval to match customers with a large product catalog through natural-language queries.
The operating problem
The product had to translate natural-language customer needs into useful matches across a catalog of more than 100,000 products. A prototype could demonstrate the idea, but production required repeatable retrieval, application integration, usable ranking context, and response times suitable for a customer workflow.
The delivered system
The production platform used Django and React around a Claude-powered retrieval experience, with Qdrant providing vector search and AWS supporting deployment. The product narrowed a large catalog into relevant candidates and presented the recommendation workflow through a maintained application rather than a standalone model demo.
Operating sequence
This sequence describes the delivered product pattern at a functional level. It does not expose confidential client implementation details.
A user describes customer criteria and constraints in natural language.
The platform converts the request into retrieval context and finds relevant catalog candidates.
The model organizes the candidate set into recommendations for the product experience.
The user reviews the result and continues the governed customer workflow in the application.
Delivered capabilities
Production controls
Figures are reported outcomes from the original portfolio material and are not presented as independently audited benchmarks.
Modern retrieval systems increasingly combine semantic and keyword signals, then rerank a smaller candidate set for precision. The broader lesson matches this delivery: useful recommendations come from the complete retrieval and product pipeline, not from the language model in isolation.
Read the primary source Qdrant: Hybrid Search with Reranking