GarbaPass

Overview

GarbaPass (garbapass.com) is a LogicSpark product: a pass exchange for Navratri 2026 in Ahmedabad, live for the nine garba nights from 11 to 19 October. Friends drop out and passes go to waste while others hunt for one across WhatsApp groups, so GarbaPass puts both sides on one live board. Sellers post for free, buyers pay a small fee to get the seller's WhatsApp number, and the support team runs every conversation from an admin WhatsApp inbox where Claude drafts the replies.

The challenge

Support for a festival marketplace happens almost entirely on WhatsApp, in a short window, and in whatever language and script each customer happens to type. Questions range from "is my post still live?" to payment lookups, and a wrong answer about a payment or a verification badge costs trust immediately. The team needed faster replies without letting an AI promise things on the business's behalf.

What we built

• A public live board with Has pass and Needs pass posts, a filter for each festival night, a Navratri 2026 calendar showing sellers and buyers per night, and a buy and sell safely guide with seven plain rules.
• Mobile and email verification badges on posts, a paid Recommended seller placement, and a paid contact unlock that reveals the seller's WhatsApp number right after payment.
• An admin console covering posts, matches, users, contacts, cities and events, partners, payments and refunds, contact unlocks and payment lookup.
• An admin WhatsApp inbox built on the Visity WhatsApp API, where each chat shows what GarbaPass already knows about that number: seller, buyer, partner or lead.
• A reply suggester on Claude through Amazon Bedrock. For the open chat it loads the recent thread (up to 20 messages), what the system knows about the person and the house-style quick replies, and returns three replies as JSON. The admin can type a rough idea first, such as "say we'll refund you", and have it written in the customer's language and script.
• Guardrails in the prompt and the route: facts only from the supplied context, a prompt-injection clause, no payment or verification claims, and admin-only access. The suggester never sends a message; a human picks, edits and sends every reply.

Results

GarbaPass runs in production on Vercel for Navratri 2026, with buyers and sellers meeting on one board instead of scattered WhatsApp groups. The support team answers WhatsApp chats from one inbox with three ready drafts per conversation, while every message still passes through a person before it reaches a customer. The same human-in-the-loop pattern, grounded context, structured output and narrow permissions, is what we use when we build AI agents for client teams. See more of our work in case studies.