AI-Powered Sales Assistant with RAG

The Challenge

A large automotive inventory-marketing platform wanted to give thousands of car dealers a single conversational assistant: one that could answer plain-English questions about their live inventory, explain pricing and risk concepts consistently, rate and rewrite vehicle listings, and improve listing photos, without ever exposing one dealer's data to another or inventing numbers.

The Solution

We built a streaming assistant on a custom, fully-typed pipeline framework rather than an off-the-shelf agent library, so every step is testable, traced, and independently evaluated. A retrieval layer grounds answers in a curated knowledge base using Amazon Titan embeddings in Postgres/pgvector, filtered by document type, tags, and user persona. An intent router dispatches each question to the right tool: retrieval-grounded Q&A, natural-language-to-SQL over BigQuery (with dealer-scoped validation and cost caps so no dealer can read another's data), live vehicle-page analysis via a resilient headless crawler, or a GPU image service that segments photos with Segment Anything, removes backgrounds, and adds feature badges using a vision model. Answers without grounding return a safe fallback instead of a guess, and every non-deterministic step is scored by an LLM judge in Langfuse plus an offline retrieval-quality suite.

Results

• Shipped to production on Kubernetes across CPU and GPU fleets with GitOps delivery
• Hard multi-tenant isolation: dealer-scoped SQL validation blocks cross-dealer data access
• Grounded RAG with an anti-hallucination fallback instead of fabricated answers
• Continuous evaluation: LLM-as-judge in Langfuse plus an offline DeepEval retrieval suite

Technologies Used