AI assistant for real estate leads on WhatsApp
Alloha explores AI-powered lead qualification, property search, and broker handoff using FastAPI, Next.js, Supabase/PostgreSQL, Redis, pgvector, WhatsApp-oriented workflows, and a guardrailed model gateway.
Status: private MVP and pilot case study. It is not a public production SaaS, and the current audit findings are disclosed below.
active listings measured in Supabase
active listings with embeddings filled during the audit
pgvector/RPC dimension represented in the repo
architecture present, not connected to primary endpoints yet
Problem
Real estate teams often lose leads outside business hours and struggle to respond quickly with relevant property options. A broker needs enough context to qualify the lead, answer property questions, and decide the next follow-up without reading a long raw conversation.
Solution
Lead qualification
Capture budget, location, bedrooms, timing, urgency, and contact intent from conversational messages.
Property search
Use the property database to prepare relevant listing suggestions and move toward semantic retrieval.
Broker handoff
Summarize the lead and candidate properties so a broker can follow up with context.
Architecture
- Next.js App Router frontend for landing, auth, onboarding, setup, dashboard, and contact flows.
- FastAPI backend with canonical /v1 chat, listing search, ingest, lead, auth, and system status endpoints.
- Supabase/PostgreSQL data model for properties, conversations, messages, leads, idempotency, and vector-ready retrieval.
- pgvector vector(384) schema, ivfflat indexes, and RPC files included for semantic search work.
- Redis-backed hooks for sessions, rate limits, idempotency, model metrics, and ingest locking.
- WhatsApp-oriented workflow design and model gateway for assisted lead conversations.
Audit Status
Honest current state
The May 15, 2026 audit found 330 active properties in Supabase and 0 active properties with embeddings filled. The repo includes pgvector/RPC 384d and RAG code, but the primary endpoints still use lexical search and filters. The main listing/chat endpoints need an API/schema contract correction before public use.
Implemented
- Real estate landing and dashboard-oriented Next.js app structure.
- FastAPI core API with chat, listings, ingest, leads, auth, onboarding, and health/status routes.
- Property ingest service with feed-first behavior, scraper fallback hooks, content hashes, and deactivation logic.
- Model gateway with hard-stop behavior, rate limits, provider metadata, and capacity-limited responses.
- Supabase schema and migrations for property data, conversations, leads, messages, and pgvector-ready retrieval.
- Benchmark script for measuring counts, endpoint latency, error rate, and vector retrieval behavior without exposing secrets.
Roadmap
- 1Correct the API/database contract for listing fields used by /v1/listings/search and /v1/chat/messages.
- 2Backfill 384-dimensional property embeddings for active listings.
- 3Connect RAG/vector retrieval to the primary search and chat flow.
- 4Add analytics for lead source, response quality, broker handoff, and search relevance.
- 5Build CRM-ready lead states, assignment, notes, and follow-up workflow.
- 6Expand voice, multimodal intake, and WhatsApp media handling after the core data path is stable.
- 7Harden multi-tenant onboarding, permissions, observability, and deployment operations.
Review the repository
The best way to evaluate Alloha is as an engineering portfolio project: read the code, inspect the architecture, and compare the roadmap against the audit findings.