Multi-Agent PropTech for the Japanese Real Estate Market
Buying property in Japan as a foreigner is genuinely difficult. You navigate 3 or 4 disconnected portals, decode legal documents in formal Japanese, and deal with a regulatory system that barely translates even with a lawyer. Platforms like SUUMO and HOME'S are search engines at heart. They help you find a listing, then leave you completely on your own from there.
The most important decision I made was treating legal compliance as a routing rule inside the AI itself, not a disclaimer buried in fine print.
Under Japan's Takken Act, an AI that gives negotiation advice or interprets contract terms without a real estate license is breaking the law. Most products handle that risk with a checkbox and move on. Instead, I built a Mediation Boundary as the very first step every user message goes through — before any other part of the system runs.
The mediation agent classifies each request into three categories: Category A (information, safe for AI), Category B (logistics like booking viewings, safe for AI), and Category C (negotiation or legal questions, must go to a licensed human). Category C requests are immediately handed off to a licensed real estate agent called 宅建士. This logic lives in the orchestrator's mediation subgraph and is reinforced through a three-tier permission model in the action engine, so no part of the system can accidentally bypass it.
The result is a product where users always know what the AI can and cannot do for them. That clarity builds trust in a way no disclaimer ever could.
A user types what they want in English or Japanese. The mediation agent reads the intent and routes it to one of six specialized AI agents covering property search, viewing scheduling, transactions, legal document analysis, renovation cost estimation, and general chat. The search agent converts natural language into a combined database and semantic search query — so a phrase like "pet-friendly 2LDK in Setagaya under 80 million yen" becomes a precise result.
A machine learning model trained on 50+ features including land value data, building age, and walking distance to stations estimates a fair market price. The entire purchase journey runs on a 7-stage state machine, so the system always knows where a buyer is and what actions are legally appropriate at that point.
The platform ships as a full monorepo: web app, iOS and Android mobile app, admin dashboard, licensed agent portal, and a LINE chatbot (LINE has 95 million users in Japan). It covers everything from property discovery through contract-ready transaction workflows, deployed on AWS in the Tokyo region for data residency compliance.
Nine specification documents on regulatory architecture, data strategy, and design decisions sit alongside the code, not as an afterthought. This project shows I can hold product strategy, AI system design, and real legal constraints in the same mental model and build something coherent enough to actually work.