May 2025 – Sept 2025 · Co-Built

SolAI: An Agentic AI Assistant for Real Estate

An agent built for real-estate workflows — property search, lead qualification, scheduling, and market analysis — backed by three-tier memory and an 87-tool orchestration layer, with every action traceable after the fact.

Results

  • Integrated Claude Flow's full 87-tool MCP registry — Gmail, Twilio, Google Calendar, and CRM among them — with intent-based routing across single, parallel, and orchestrated execution strategies
  • Three-tier memory (Redis, Supabase, Pinecone) enabling context recall across conversations far longer than any single prompt window
  • Security sandbox with threat detection (injection, escalation, exfiltration) and a cryptographic audit-hash chain, so every tool call the agent makes is traceable after the fact
  • Appointment-safety lifecycle (request → agent response → lead confirmation → final confirmation) built specifically so the agent can't overcommit someone's calendar

Stack

  • Node.js
  • Express
  • Claude Flow (MCP)
  • Redis
  • Supabase
  • Pinecone
  • OpenRouter (Gemini 2.5 Flash, Claude 3.5 Haiku)
  • WebSocket

Step 1

Every channel, one place

Telegram, Gmail, Instagram, Facebook, Supabase, Google Drive, Google Calendar — the dashboard surfaces every integration the agent can actually act through, so it's clear what "the agent can send an email" or "the agent can check the calendar" really means underneath.

Step 2

The assistant introduces itself and states what it can do

AirWrecka — the assistant's real-estate persona — opens by naming exactly what it helps with: property search, lead qualification, market analysis. No generic "how can I help you today" that leaves the user guessing at capability.

Step 3

Text or voice — same assistant

The same reasoning and tool access sit behind either input mode. Quick-action shortcuts (Search Database, Calendar, Send an Email, Web Search) map directly onto the tool-orchestration layer, giving a fast path into the specific integration a user actually needs.

Step 4

Operators still see what's running

Workflow run history and a notification feed sit alongside the chat, so a human overseeing the agent can see what it has actually done — not just talk to it blind and hope.

InteractiveMemory tier trace

Select a query above to trace it.

Pick a sample query to see which memory tiers it hits, in what order, and what each returns.

← Back to all work