mahimairaja/realtyrecall

RealtyRecall is a voice assistant built specifically for solo real estate agents. It answers incoming calls on behalf of the agent, qualifies buyers by asking about their budget, timeline, and financing, and books property showings against the agent's real calendar. When a call ends, it texts the lead summary to the agent's phone number. Setup starts by pasting the agent's website URL. The system crawls the site, extracts property listings and the agent's name, agency, and tone from the page content, and stages everything for the agent to review before going live. The assistant then answers all calls in that agent's name and persona, using only the listings that were found on that website, never inventing details. The standout feature is persistent memory across calls. When someone calls for the second time and provides their phone number or name, the assistant recognizes them, greets them by name, and picks up the conversation from where the previous call left off. This memory is stored in a knowledge graph using a tool called Cognee, which keeps track of which buyers expressed interest in which homes. A caller can also ask to be forgotten, which deletes all their stored data. The system uses LiveKit to handle real-time voice calls. On a web call, house cards appear on screen as the assistant mentions homes during the conversation. Bookings are created via the Cal.com calendar API, with idempotency checks to prevent double-bookings if a request is retried. A one-line JavaScript snippet lets agents add a "Talk to us" button to their own website, and a SIP phone number can connect real incoming calls to the same assistant. The backend is built with FastAPI. The assistant model used is GPT-4.1-mini, with Deepgram for speech-to-text and Cartesia for text-to-speech. Memory storage combines a Neo4j graph database with pgvector for vector search. The project was built for a hackathon and is MIT licensed. The full README is longer than what was shown.