MURENA, a MUlti-Agent LLM pipeline for large-scale Real Estate maNAgement

De Luca, M., Monaco, S., Tcaciuc, C., Gorrino, A., De Luca, G., Apiletti, D., & Cerquitelli, T.
Applied Data Science Track – ECML PKDD
2026

Managing large-scale public real estate portfolios requires screening thousands of buildings against technical, energy, spatial, and regulatory constraints; a process that currently takes domain experts weeks of manual cross-referencing across fragmented data sources.

We present MURENA, a multi-agent Human-in-the-Loop platform where the LLM acts as a semantic controller, translating user intent into structured constraints and orchestrating execution across SQL and geospatial engines, while the final decision remains with a domain expert. 

Role-specialised agents (energy performance, zoning compliance, spatial accessibility, regulatory constraints) operate over a unified registry of 90,000 Italian Energy Performance Certificates enriched with geocoded proximity indicators and an expert-curated regulatory knowledge base, through three phases: requirements extraction, SQL query composition, and multi-criteria scoring, with full intermediate traceability.

Evaluation combining automatic benchmarking on 486 queries and expert validation on 10 institutional change-of-use scenarios shows 90.3\% perfect agent routing, 90\% top-5 retrieval of expert-preferred buildings, and end-to-end execution in tens of seconds, reducing manual screening from weeks to minutes.

The platform, developed with institutional stakeholders and running on locally hosted open-weight models, has attracted interest from regional and national bodies for operationalising building decarbonisation under EPBD IV.

[De Luca et al., 2026] De Luca, M., Monaco, S., Tcaciuc, C., Gorrino, A., De Luca, G., Apiletti, D., & Cerquitelli, T. (2026). MURENA, a MUlti-Agent LLM pipeline for large-scale Real Estate maNAgement. Applied Data Science Track – ECML PKDD 2026, Naples, Italy, September 7–11, 2026. Accepted for publication.