Description
Summary AI Voice Receptionist — MVP Build Brief Project Summary Build a single AI voice agent that answers inbound calls for small local businesses (initial vertical: mechanics/HVAC), looks up business-specific info via RAG, answers customer questions, books appointments, and escalates emergencies — 24/7. One agent architecture serves multiple businesses via per-business knowledge packs, not a separate bot per client. Goal: A working demo-ready MVP that can take a real inbound call for any one of ~10 pilot businesses, answer correctly using that business’s info, book a slot on a calendar, and text/forward true emergencies to the owner. Scope (v1 — lean, not polished) In scope 1. Inbound call handling — dedicated phone number(s) receive calls, route to the AI agent. 2. Speech-to-text → LLM → text-to-speech pipeline — real-time conversational voice. 3. Per-business knowledge lookup (RAG) — each business has its own data (services, hours, pricing, FAQs) stored as embeddings; the agent retrieves the right business’s context per call (call routed by the number dialed, or a business ID passed in). 4. Appointment booking — agent checks availability and books directly into a shared calendar (Google Calendar). 5. Emergency detection & escalation — keyword/intent detection (e.g. “no heat,” “gas smell,” “car won’t start on the highway”) triggers an SMS or call forward to the business owner’s cell in real time. 6. Simple data ingestion — a way (form, spreadsheet, or scraped-and-pasted text) to load a new business’s info into the knowledge base. Does not need an admin UI — manual/CLI/script is fine for v1. 7. Call logging — every call transcribed and stored (for QA and for showing prospects “here’s what it heard/said”). (Note: CRM integration will happen later per business) Call Flow (minimum viable) 1. Call comes in on a business’s assigned number. 2. STT transcribes caller speech in real time. 3. System identifies which business this call belongs to (by number dialed). 4. RAG query pulls relevant chunks from that business’s Supabase namespace. 5. LLM generates a response grounded in that context, speaks it via TTS. 6. If caller wants to book: agent checks calendar availability, confirms a slot, writes the event. 7. If caller’s message matches an emergency pattern: agent tells caller help is being alerted, and simultaneously fires an SMS/forward to the owner. 8. Full transcript + outcome (booked / info given / escalated) logged. Deliverables 1. Working backend that handles the call flow above end-to-end for at least one live test business. 2. Script/process to onboard a new business (add its info to the knowledge base) in under 15 minutes. 3. Brief technical README: architecture diagram, how to add a business, how to point a new phone number at the system, environment/config setup. 4. Deployed and reachable via a real phone number for live demo calls. Acceptance Criteria • A real call to the test number is answered by the AI within 2 rings. • Agent correctly answers at least 5 out of 5 pre-scripted test questions using that business’s specific data (not generic answers). • Agent successfully books a test appointment that appears correctly on the calendar. • A scripted “emergency” phrase triggers an SMS to a test phone number within 10 seconds. • Full call transcript is retrievable after the call. What I’m Providing • List of 10 pilot businesses (name, services, hours, pricing — will supply as onboarding data) • API keys/accounts for chosen third-party tools (or freelancer can recommend and I’ll provision) Provide 1. Sample 2. Budget estimate 3. Proposed stack 4 proposed duration to compete 5. How quickly you can start Generic applications will not be considered