AI Lead Conversion System for Real Estate

The Problem

Disconnected CRM tools create leakage between first touch and booked appointment and make pipeline consistency hard to enforce.

The Solution

Built a multi-tenant conversion platform that unifies conversational capture, AI qualification, deal lifecycle management, and analytics. Enforced tenant isolation and deterministic stage transitions in Supabase-backed APIs, with optional async delivery via FastAPI and Redis.

Impact

Enterprise-grade conversion workflow with tenant isolation and reliable stage transitions.

Architecture

Key Challenges

• Multi-tenant data isolation was the hardest architectural constraint. A misconfigured RLS policy could expose one agency's leads to another. Wrote a test suite that impersonates different tenant JWTs and asserts that cross-tenant queries return empty results — ran this on every schema migration.
• Deal stage transitions needed to be deterministic — agents couldn't skip stages or move backwards without an explicit override. Implemented a finite state machine in TypeScript that validates every transition against an allowed-transitions map before writing to Supabase.
• FastAPI and Next.js needed to share Zod validation schemas without duplication. Extracted all shared schemas into a single TypeScript package imported by both the Next.js app and a Pydantic-compatible auto-generated Python schema, keeping both layers in sync.

Key Learnings

Multi-tenant architecture decisions made early are very hard to change later — RLS policies, schema namespacing, and JWT claim structure all need to be designed together from day one, not bolted on. I also learned that Redis as an async queue between FastAPI and Supabase adds meaningful complexity for modest throughput gains — for most real estate agency scales, a simpler Supabase edge function would have been sufficient.