Back to case studiesAI / Automation

WhatsApp AI Support Bot — Real Answers, Real Human Handoff

A WhatsApp customer-support bot that answers real questions from a real FAQ knowledge base, transcribes voice notes, and deterministically escalates anything it can't confidently handle to a real person by email — live and testable right now.

FlaskGoogle Gemini (genai)Twilio WhatsAppUpstash RedisVercel

Live on Vercel — verified end-to-end with a real Gemini API key

View Live Demo View Source

The Problem

Small businesses lose leads answering the same WhatsApp questions on repeat, and a bot that guesses wrong (or can't tell when to hand off) costs more trust than it saves.

Our Approach

1Built a classify-then-route pipeline: Gemini labels each message (greeting, FAQ, or complex), with a deterministic keyword/word-boundary override so a refund request or the word "HUMAN" always escalates, regardless of what the classifier says.
2Every module fails toward safety, not silence — no Gemini key configured means the bot still replies and escalates instead of crashing or guessing, verified by testing the entire flow with zero credentials before a single real API call was made.
3Deployed on Vercel's native Python runtime (no legacy serverless-wsgi wrapper needed), with Upstash Redis for persistent memory and an admin dashboard showing real conversation stats — no fabricated "uptime" for what is a stateless deployment.

What This Achieves

FAQ questions get answered instantly from a real knowledge base — the bot is instructed to say "not sure" and escalate rather than invent an answer.
Anything the bot can't confidently handle reaches a real person by email within seconds, with the recent conversation already attached.
The whole system runs on Vercel's free tier and degrades gracefully at every layer — a missing credential never means a broken bot.