A production-ready serverless re-architecture featuring an AI concierge and real-time operational staffing.
Last updated: May 21, 2026
Gamlebyen Mat & Drikke AS (St. Halvards gate 33, 0192 Oslo) operates Kanon Catering in Oslo, Norway, handling high-effort manual operations: slow manual quotations outside business hours, and friction coordinating service staff shifts in real time. These manual workflows create bottlenecks — quotes pile up unanswered overnight, and staffing changes require phone chains and spreadsheets that drift out of sync. The business needed a modern platform to centralize operations, automate lead capture, and give both admins and workers a single source of truth.
A production-ready, serverless re-architecture built around Convex as the real-time backend. Every operational action — quote requests, staffing updates, admin approvals — flows through a strongly-typed, event-sourced schema. The frontend is Astro 5 with selective React 19 islands for interactive surfaces, deployed to Cloudflare Workers for edge performance. Better Auth with RBAC protects admin and worker routes. The AI concierge, powered by Google Vertex AI, converts natural-language inquiries into structured leads without human intervention. Email delivery is handled by Resend with SMTP fallback via Nodemailer. The result: a fully functional staging platform that demonstrates next-generation architectural capability without disrupting the client's existing legacy systems.
How it works
Customer describes their event via chat. Vertex AI extracts date, guests, dietary needs, and service type into a structured lead.
Administrators review the AI-generated quote, adjust pricing, and approve. Resend dispatches the finalized quote to the customer instantly.
Workers confirm shifts via the portal. Convex pushes schedule changes in real time to both admins and workers.
The platform is organized around three technical pillars that eliminate manual work and create real-time operational visibility.
A conversational agent built on @convex-dev/agent and Google Vertex AI (gemini-2.5-flash-lite) that captures customer intent through natural chat. It extracts structured data — event date, guest count, dietary restrictions, service type — and writes qualified leads directly into Convex. The concierge operates 24/7, eliminating the quote backlog that builds up outside business hours. No human operator is required for initial qualification.
A protected admin interface where administrators review AI chat transcripts, edit generated quotes, and calculate taxes in real time. The dashboard surfaces lead status, event timelines, and staffing assignments in a single view. Automated email notifications are dispatched via Resend when quotes are approved or updated, with SMTP fallback ensuring delivery resilience.
A dedicated worker portal at /worker where staff confirm shifts, sign attendance via geolocation, and synchronize their schedules with the admin calendar. The portal is role-gated through Better Auth with separate admin and worker RBAC roles. Real-time updates flow through Convex so schedule changes propagate instantly to both admins and workers.
The system is maintained as a fully functional technical demonstration environment on Cloudflare Workers. This staging approach serves a specific engineering purpose: it showcases the complete architectural stack — real-time backend, AI orchestration, RBAC-protected routes, and edge deployment — without interfering with the client's existing production website and business workflows. The staging environment proves that the platform is production-ready and can be migrated to the client's live domain when they choose to transition. It is a deliberate engineering transparency choice, not a limitation.
Key outcomes and metrics from the implementation. Every number is drawn from the project scope, the staging system, or authoritative industry data.
A couple requests a wedding quote at 23:45. The concierge extracts date, guests, dietary needs, and service type, then generates a structured lead before the office opens. Zero human intervention.
Norwegian and English served from the same Astro build. One content schema drives both locales. Language switches update reactively through Convex — no page reload.
A worker calls in sick at 06:00. The admin updates the shift. Within milliseconds, every affected worker's phone updates via Convex sync. No phone chains. No missed shifts.
AI chatbots for lead qualification reduce average handling time by ~35% (Marketing LTB, 2026). The Kanon concierge qualifies overnight requests before the first staff member checks email.
Have a project in mind? I'd love to hear about it. Whether you need an AI agent, a full-stack platform, or something that doesn't exist yet — let's build it.