Brief
A three-module AI platform for a marketplace serving over 200,000 people a day. One retrieval layer powers customer support, on-platform listing insights, and personalised recommendations.
+10%
GMV Lift
−40%
Support Tickets
200k
Daily Users
Problem
Support volume was growing faster than the team could hire. Sellers were writing listings that did not convert, and buyers could not surface the right one among thousands. Each of those was being solved manually, and none of them scaled with traffic.
Solution
- A multi-tenant chatbot with persistent memory, answering from the platform knowledge base through retrieval rather than a general-purpose model.
- An SEO analyser that scores listings in real time and suggests rewrites, so sellers fix conversion problems before they cost them a sale.
- A vector-based recommendation engine returning personalised results in under 100ms at full daily load.
- A cost-optimised inference pipeline, using embedding caching plus model routing, so none of the above scaled the bill linearly with usage.
Result
GMV rose 10% and support tickets fell 40%. Listing conversion improved 3.4× where sellers acted on the analyser, and the caching and routing work cut LLM spend by 62% against the naive implementation.
Under the hood
| Backend | FastAPI, Redis, AWS Lambda |
|---|---|
| AI / retrieval | OpenAI, Pinecone vector search, embedding cache, model routing |
| Frontend | React |
| Scale | 200k+ daily users, sub-100ms recommendation latency |
