Restaurant Voice Ordering & Operations Copilot

Restaurants lose revenue and guest satisfaction when ordering is friction-heavy — long hold times, menu confusion, modifier mistakes, and allergy-related risk. Traditional IVR and static web forms can't handle natural conversation ("two medium-spice butter chickens, one without dairy, pickup in twenty minutes") while still enforcing business rules, and generative AI alone isn't safe for food service — models can invent ingredients, skip allergen warnings, or confirm orders that were never validated.

Build a voice- and chat-first ordering copilot, piloted for a real Indian kitchen, that feels natural to guests while treating the database — not the model — as the source of truth, across multiple trust boundaries: conversational AI at scale, safety-critical allergen handling, a hybrid architecture pairing LLM dialogue with server-side enforcement, and a full staff operations loop.

• The agent must never write directly to the database — every menu fact, cart change, and confirmation flows through validated server tools
• Allergen data is classified (contains / may contain / cross-contact / unknown) and the system must never claim medical safety
• Production hardening: fail-closed webhook authentication, CORS allowlists, JWT-scoped staff access, idempotent ticket dispatch
• Voice agents miss checkout phrases, skip tool calls, or duplicate responses without added reliability layers

Guests browse the menu, build a cart, receive allergen warnings, confirm orders with a full read-back, and trigger kitchen tickets by voice or text in a single session. ElevenLabs Conversational AI handles natural dialogue over WebSocket text and WebRTC voice; Supabase holds menu, orders, sessions, escalations, and the audit trail; Edge Functions enforce validation, cart mutations, confirmation, and ticket dispatch; n8n orchestrates kitchen/POS workflow; and a React monorepo serves both the guest ordering embed and a staff operations dashboard.

• ElevenLabs Conversational AI for natural dialogue (WebSocket text + WebRTC voice)
• 13 ElevenLabs agent tools for menu search, cart, allergens, read-back, confirm, and escalate
• 14 Supabase Edge Functions and 53 shared server modules enforcing validation, cart mutations, confirmation, and ticket dispatch
• n8n for kitchen/POS workflow orchestration
• React monorepo serving the guest ordering embed and staff operations dashboard
• Client-side reliability bridges (checkout phrase detection, escalation, session wrap-up) pairing with server-side enforcement (confirmation gates, rate limits, webhook secrets, hashed session tokens)

• The agent never writes to the database — every menu fact and cart change flows through validated server tools, with explicit guest approval before a kitchen ticket exists
• When allergen data is incomplete the system warns and escalates to staff rather than guessing; safety copy is templated from backend classification, never model-generated
• Split conversation from commitment — the LLM handles dialogue, the backend enforces state transitions for order confirmation and ticket creation
• Added client-side reliability bridges for checkout, escalation, and session wrap-up to cover gaps in even well-prompted voice agents

• Ground the agent in tools, not memory — menu prices, ingredients, and availability must come from structured server-tool data, or hallucination follows
• Split conversation from commitment — the LLM handles dialogue; the backend handles state transitions
• Voice agents need reliability layers — client-side phrase detection and deterministic bridges materially improve pilot stability without sacrificing natural conversation
• Regulated domains need explicit uncertainty handling — when allergen data is incomplete, warn and escalate rather than guess
• Modular architecture de-risks iteration — voice provider, database, workflow engine, and frontend can evolve independently
• Auditability is a feature, not overhead — every tool invocation and safety-relevant decision is logged for operators and future compliance review