Self-Improving Real Estate Valuer

Inspiration

I moved to Dubai late last year, and Al Jadaf opened my eyes to the incredible complexity of the local real estate market. Standard search tools often feel more like estimates than factual figures, and they also fail to account for the rapid shifts in community trends or the nuances between established areas and emerging developments. I wanted to build a system that doesn't just calculate a price once but actively identifies its mistakes and updates its "internal manual" to get smarter with every transaction it reviews.

What it does

How we built it

The project is architected as a multi-agent system using LangGraph to manage a persistent state across runs. I designed a reflective loop where three specialised models act as a team:

The Valuer: Acts as the primary professional, generating valuations based on current "Strategy Logs" and historical benchmarks.

The Auditor: Uses deep reasoning to compare the Valuer’s output against the actual market worth. It performs a root-cause analysis to see if the valuer missed a systemic trend, like a premium for off-plan units.

The Strategist: Acts as the "Learning Engine". It takes the auditor’s critique and updates a global strategy log, a shared memory that dictates how the valuer should think in the next round.

Challenges we ran into

The biggest hurdle was state persistence. In typical LLM applications, the model "forgets" the feedback once the conversation ends. By using LangGraph, we had to ensure the strategy_log was passed back into the system prompt for every new query. Another challenge was data orchestration. Merging real-time GeoJSON spatial data with Metro CSVs to give the agents a "sense of place" so they could understand how proximity to infrastructure impacts value

Built With

• claude
• langgraph