Hospitality Concierge Platform
One WhatsApp concierge engine, many businesses — per-client knowledge with fail-closed isolation.
Problem
Hostels, bars, and tour operators answer the same guest questions all day in WhatsApp groups — but a shared bot that could leak one business's private data into another's chat is worse than no bot.
Solution
A multi-client message engine — every group resolves to exactly one client or the message goes nowhere; replies are grounded in that client's knowledge; every reply is stored with an audit trail. One engine, many deployments.
Architecture
The concierge is a platform bet: hospitality businesses in tourist destinations share the same question patterns, so one engine over different data serves them all. The engineering priority was isolation before intelligence — group-to-client resolution fails closed, private knowledge is scoped per client, and the public destination layer is deliberately deferred until the private core is proven.
Repo evidence
- End-to-end pipeline built — ingest → trigger gate → context → LLM → stored reply with audit; MVP schema applied to a live dev database
- Transport-only container built, session-persistent, ready for a persistent host
- Multi-provider LLM registry (OpenAI / Anthropic / Google / xAI)
- Written retrieval upgrade path — v1 is deliberately retrieval-free until corpus size justifies vectors