AI Auction Assistant

The problem

A vehicle auction platform was hitting the ceiling of its synchronous architecture. Under bidding load, latency blocked the user experience. Listing data sat in silos, so staff compiled vehicle summaries by hand. And the platform's recommendations were a black box, which is a real problem when buyers are committing five-figure sums on trust.

The build

An AI assistant embedded directly into the auction flow, built on an event-driven architecture with four layers. A WebSocket layer keeps every client in sync at 60-100ms, even during bidding spikes. A Redis-backed task queue absorbs the heavy work asynchronously, so nothing blocks the auction itself. A vector database stores behavioural data and clusters buyers by their behaviour. On top sit six AI modules: vehicle summaries, damage assessment, recommendations, conversational search, staff engagement tools, and cost tracking.

The design decision that mattered most: recommendations are explainable. When the system suggests a vehicle, it shows why - the behavioural signals and listing attributes behind the match. Buyers can disagree with the reasoning. They could never disagree with a black box.

The workflow

Every user action lands as an event and fans out through the queue. AI work happens off the critical path and streams back over the socket when ready. The LLM layer is model-agnostic, so the platform can swap providers as economics change, and a cost-tracking module ties AI spend to the bid volume and sell-through it generates.

The outcome

Staff stopped writing vehicle summaries by hand. Buyers got recommendations they could interrogate. And the platform owner got the dashboard every AI project should have: what the models cost, and what they return.