Inspiration
The United States had approximately 5.9 million commercial buildings covering 96.4 billion square feet in 2018, according to the U.S. Energy Information Administration (EIA).
Yet despite this enormous scale, the EIA identifies a fundamental challenge: no comprehensive national inventory of commercial buildings exists.
Even the Lawrence Berkeley National Laboratory's Building Performance Database, containing information on over a million buildings, reports that detailed operational and building-system information is available for relatively few properties.
The problem isn't simply a lack of data. It's that valuable information about the built environment is fragmented across property records, aerial imagery, building models, and disconnected databases, making meaningful research difficult, time-consuming, and challenging to verify.
Answering questions such as "Do older office buildings dedicate more land to surface parking?" can require gathering multiple datasets, reconciling inconsistent information, identifying missing measurements, and performing statistical analysis.
We asked ourselves:
What if AI could conduct the research instead of simply generating an answer?
That question inspired UrbanLab — An AI Scientist for the Built Environment.
Rather than building another chatbot that summarizes information, we wanted to create a system that could plan investigations, gather evidence, analyze data, and produce transparent, reproducible findings.
What it does
UrbanLab transforms natural-language questions about the built environment into structured, evidence-driven investigations.
For example, a researcher might ask:
"Do older office buildings in Houston dedicate more land to surface parking?"
Instead of immediately generating an answer, UrbanLab follows a research workflow:
Question → Research Plan → Evidence Collection → Dataset Construction → Statistical Analysis → Findings
The system interprets the research question, identifies relevant variables, develops an investigation plan, and presents it for human approval before execution.
Once approved, UrbanLab coordinates specialized research agents and data tools to:
- Explore commercial properties: Query and filter property records based on location, building characteristics, and research requirements.
- Collect evidence: Retrieve relevant property information and incorporate available spatial evidence.
- Construct research datasets: Organize measurements into structured datasets with traceable sources.
- Identify evidence gaps: Explicitly flag unavailable or incomplete information rather than inventing values.
- Perform statistical analysis: Investigate relationships, compare groups, and evaluate uncertainty using programmatic statistical methods.
- Present findings: Display research progress, property information, maps, datasets, visualizations, and conclusions through an interactive dashboard.
UrbanLab currently focuses on commercial-property research in Houston and Los Angeles, with an architecture designed to support additional locations and research questions.
Our guiding principle is:
No claim without evidence.
The AI determines what needs to be investigated, while property values and numerical computations come from dedicated data and analysis tools rather than being generated by the language model.
How we built it
We built UrbanLab using Next.js, React, TypeScript, and Tailwind CSS for the frontend, with a Python and FastAPI backend.
Our research workflow uses Google Gemini and LangChain DeepAgents, where a principal investigator coordinates five specialized agents for research planning, data collection, visual measurement, gap detection, and analysis. Users review and approve research plans before execution.
We integrated CoStar's commercial-property dataset for building information and developed support for Vexcel aerial imagery and Matterport 3D model references.
For reliability, property searches use structured filters, numerical analysis is performed using Python rather than generated by the AI, and results retain their data sources. Missing information is explicitly identified instead of being fabricated.
Challenges we ran into
Making autonomous research trustworthy
Our biggest challenge was ensuring that UrbanLab behaved like a research system rather than a chatbot that simply generated convincing answers.
We needed to preserve evidence sources, avoid fabricated property values, maintain reproducible calculations, and communicate uncertainty clearly.
Integrating heterogeneous datasets
Commercial-property records, aerial imagery, and building models contain different schemas, identifiers, units, and levels of detail.
Connecting these sources required careful normalization, property matching, field mapping, and validation.
Handling incomplete evidence
Not every research question can be fully answered with the available information.
For example, a property record may report parking-space counts without providing the physical area occupied by surface parking.
We had to distinguish reported measurements from estimates and ensure that missing information was communicated rather than silently filled in.
Coordinating multiple AI agents
Building an autonomous research workflow required us to define clear responsibilities for each agent, coordinate their outputs, maintain consistent data structures, and introduce human oversight.
We learned that adding more agents or data sources doesn't automatically improve research. The system needs to select the right tools and evidence for each question.
Responsible sponsor-data handling
Working with sponsor-provided datasets and external APIs introduced additional engineering challenges involving credential security, controlled API usage, data access restrictions, and reproducibility.
Accomplishments that we're proud of
- Built a multi-agent AI research system that transforms natural-language questions into structured investigations.
- Integrated real CoStar property data with filtering, analysis, and evidence tracking.
- Developed an interactive dashboard for exploring properties, research progress, and findings.
- Implemented human-approved research plans and transparent handling of missing data.
- Prototyped Track This Study using simulated historical data to visualize how findings evolve over time.
We're proud of creating a system that goes beyond AI-generated answers to deliver evidence-backed research.
What we learned
Building UrbanLab changed how we think about applying AI to real-world research.
The most important question isn't:
"Can an LLM generate an answer?"
It's:
"Can an AI system determine what evidence is needed, collect it responsibly, and produce a conclusion that can be verified?"
We learned that autonomous agents are most useful when they coordinate reliable tools rather than replace them.
We also discovered that missing information is itself an important research finding. A trustworthy system must be able to explain what it knows, what it estimates, and what it cannot determine.
Finally, we learned the importance of combining AI reasoning with deterministic data processing, statistical validation, and human oversight.
What's next for UrbanLab
Our vision is to evolve UrbanLab from an AI-powered research platform into a continuous research system that can investigate how the built environment changes over time.
One feature we're exploring is Track This Study, which would allow researchers to revisit completed investigations and monitor how their findings evolve.
We developed an initial demonstration using simulated historical data, illustrating how a study could be tracked across multiple runs and how changes in key metrics, confidence intervals, and research conclusions could be visualized.
In the future, we want to extend this into a fully automated longitudinal research workflow that can:
- Automatically revisit research questions: Rerun studies on weekly, monthly, or quarterly schedules using refreshed evidence.
- Detect meaningful changes: Compare new findings against previous results and distinguish statistically significant changes from normal variation.
- Track evolving conclusions: Identify whether a finding remains unchanged, strengthens, weakens, or requires revision.
- Integrate Cartesia voice AI: Conduct consent-based interviews with property stakeholders to gather information unavailable in structured datasets, such as parking arrangements, building operations, and infrastructure readiness.
- Expand spatial intelligence: Further validate live Vexcel aerial measurements and Matterport 3D building context.
- Scale research coverage: Support additional cities, property types, and research questions.
- Improve reproducibility: Expand research exports, evidence documentation, and longitudinal reporting.
Our long-term goal is to build an AI scientist that doesn't just answer a research question once, but continues investigating it as the world changes.
UrbanLab isn't just about asking AI questions about buildings. It's about giving AI the tools to investigate those questions scientifically.
