EventQuote AI

Event vendors spend hours assembling quotes manually — pulling specs from PDFs, checking vendor availability, formatting line items. The process is inconsistent enough to lose deals to competitors who respond faster. EventQuote AI replaces that workflow with a retrieval-augmented pipeline that reads a client brief and generates a structured, catalog-grounded quote in under two minutes.

What I built

The core is a three-stage RAG pipeline. Ingestion: vendor rate sheets and product catalogs are chunked, embedded, and stored in PostgreSQL with pgvector; a background worker keeps embeddings current as vendors update pricing. Retrieval: at query time the client brief is embedded, top-k chunks are fetched by cosine similarity, and a reranking pass filters noise before generation.

FastAPI serves the API layer. Redis caches embedding lookups for repeat queries. Everything runs in Docker on Cloudflare Workers backed by a managed Postgres instance.

Hard parts

Catalog freshness vs. query latency. Vendors update pricing frequently, so stale embeddings produce wrong quotes. The naive fix — re-embed on every update — blocked reads. Decoupling ingestion into an async worker with a write-ahead log solved this: the query path always reads from a consistent snapshot while the worker catches up.

Hallucination containment. LLMs will confidently invent plausible-sounding line items. A post-generation validation pass checks every quoted item against the retrieval context; anything ungrounded gets quarantined. This added ~30ms to p99 latency but eliminated the category of error that most eroded vendor trust.

Result

24k requests/day at p99 180ms end-to-end. 99.8% uptime over 90 days in production. The retrieval-quality floor from hybrid search reduced quote revision requests by roughly a third compared to dense-only retrieval in A/B testing.