Rebound

Rebound helps field-service teams reschedule appointments from customer messages without handing booking authority to a language model. The problem is simple: HTTP 200 proves an API accepted a request, not that the customer authorized the result. An agent can misunderstand “6 pm only,” book a morning slot, and confidently report success. Outages and uncertain writes make these mistakes harder to detect. Rebound’s Resolve service uses Meta Muse Spark to extract typed constraints from customer messages, booking history, and technician information. Deterministic code validates the intent, searches availability, and requests clarification or approval. In confirmation mode, the customer approves the exact slot before booking. The model never writes bookings directly. A PostgreSQL-backed queue, versioned approvals, worker leases, and idempotency keys support outage recovery and reconciliation after worker crashes. SMTP delivers customer notifications, CalDAV mirrors bookings to technician calendars, and an independent SQL oracle checks the actual database outcome. The repository includes a working local application, frozen synthetic evaluation cases, recovery checks, and reproducible setup instructions. It also preserves the fault-injection study that motivated the product: evaluating what agents actually changed, rather than trusting what they said.