Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Robert Chun

Second Advisor

Suneuy Kim

Third Advisor

Kavindra Yerolkar

Keywords

Geospatial AI, Large Language Models, Agentic Systems, Spatial Reasoning, Urban Intelligence, ReAct Framework

Abstract

Decision making in urban real estate domain requires use of heterogeneous spatial parameters like parcel records, zoning overlays, transit infrastructure, schools, permit histories, and demographic statistics. The existing geospatial tools give access to these parameters, helping the Geographic Information System (GIS) experts to interpret insights out of a large set of parameters using complex tooling. However, interpreting value out of the data requires reasoning and geospatial domain knowledge. The current state-of-the-art language models can provide abstraction over these complex analytical tools by allowing a homebuyer or urban planner expert to query in natural language to retrieve information. However, bare Large Language Models (LLM) responses to geospatial queries sometimes exhibit spatial inaccuracy, producing plausible sounding but spatially hallucinated answers. Even with factually correct responses, it is hard to consume this information and reason across multiple spatial parameters like school ratings, crime, transit, and flood risk presented in plain text. This thesis presents MediniAI, a spatially grounded agentic AI system with reasoning capability and domain knowledge, by providing tools to ground the response and help visualize using MapStoryboard. MediniAI provides a cinematic sequence of georeferenced map scenes with coordinated narrative and animated camera paths, which are rendered in a React and MapLibre GL frontend with real-time server-sent event streaming. MediniAI uses twelve domain-specific tools callable by the reasoning agent. The system implements two architectural patterns: a single-agent Reasoning and Acting (ReAct) loop bounded at eight reasoning turns, and a multi-agent approach composing of orchestrator agent, a data-retrieval agent, and a narration agent. A detailed comparison of agentic operations using LLMs (OpenAI GPT-4o, Anthropic Claude Sonnet 4.6, and Google Gemini 3.1 Pro) and seven structured experiments evaluate the central hypothesis that spatially grounded agentic reasoning achieves significantly higher spatial accuracy than bare LLM. For factual queries the results were evaluated against ground truths backed by a database of approximately 1.3 million rows — spanning parcels, businesses, transit stops, schools, census tracts, zoning overlays, permits, crime data, and rent indices. Geometric verification (Haversine distance, iv point-in-polygon) was used for spatial claims and LLM-as-judge for open-ended queries. The statistical significance was assessed via Mann-Whitney U tests across experimental conditions.

Download

Available for download on Saturday, May 22, 2027

Share

COinS