Hospitality Agent
A back-office AI for a cafe and events operation — building the crew rota, coordinating suppliers, costing food and drink, triaging customer enquiries, and reporting weekly.
Rota as a constraint problem
Crew scheduling is solved against real constraints — availability, hours, skills, cover — rather than hand-juggled, then drafted for a human to confirm.
The conductor + task-agent shape
A conductor classifies incoming work and routes it to focused task agents (mail, calendar, files, voice). High-volume classification runs on a local model to keep cost flat; the heavier reasoning uses the shared subscription bridge.
Reports that synthesise, not dump
A weekly engagement report is synthesised from the underlying data into something readable, rather than a wall of raw figures.
Local model for the cheap, heavy lifting.
Routing and classification — high volume, low difficulty — run on a local model, reserving the foundation model for work that needs it.
Draft, then confirm.
The agent prepares rotas, replies and orders; a person approves before anything is sent.
What’s open source
Built on the open-source Pandora’s Box platform and its public capability modules (conductor + the mail / calendar / files / voice task-agent modules). The tenant-specific business logic and content for this project are private — this page documents the methods, not a public drop of the project itself.