Q.Promo AI — Promotion questions answered in seconds, not days

The problem

Retailers and suppliers run promotions worth millions, but when someone asked “how did last month’s promo perform in this category?”, the answer took days. An analyst had to pick up the request, write custom SQL, and send back a report — and two analysts would often give two slightly different answers to the same question. Decisions waited on a queue.

What I built

A chat interface where anyone — a category manager, a supplier, someone with no SQL knowledge — asks the question in plain English and gets an answer they can trust.

The core idea: the system knows when to be certain and when to think. Questions it has seen before go down a pre-validated path and come back in 10–12 seconds with a guaranteed-correct answer. New or unusual questions go to an AI reasoning path that writes and verifies its own SQL, taking 20–30 seconds. And when a novel question is answered correctly, the system saves it — so tomorrow it’s an instant answer.

What changed

• –Analysis that took days now takes minutes end-to-end; 90% reduction in turnaround time
• –70% of questions get instant, pre-validated answers
• –Roughly half of the analyst team’s capacity was freed for work that actually needs a human
• –Stakeholders got self-service access around the clock, with every answer traceable

Under the hood

LangGraph multi-agent workflow (Gatekeeper, Orchestrator, SQL Verifier, Narrative Generator) on Google Vertex AI Gemini. Qdrant vector store for validated playbooks, a NetworkX knowledge graph for database schema traversal, BigQuery underneath, Next.js 15 frontend integrated into Google Chat, full observability via Langfuse. Docker Compose on GCP. Team of 4; I was the architect.