BERL is partnering with NYSERDA to aid in identifying desired use cases for Generative AI (GenAI) for Commercial Real Estate. Unlike traditional AI systems that primarily recognize or classify information, GenAI can generate original material that resembles human creativity and reasoning. Its abilities include producing coherent and contextually appropriate writing or artwork, summarizing complex information, understanding language and intent, writing and explaining computer code, analyzing data, and adapting outputs to a user’s tone or goals.
There is a large untapped potential for the use of this technology in Commercial Real Estate to understand large, disparate and unstructured data sets about building performance and to translate those insights into actionable knowledge for decision makers.
BERL has been working with a group of New York City-based Real Estate Professionals to create GenAI prototypes from Chatbots to interact with NYC Open data, to risk assessment reports to identify Local Law 97 exposure, and open DOB violations.
Our prototypes are available below, with videos showcasing their capabilities. If you’d like to test one, sign in/up for hugging face (a free GitHub-like service that hosts our websites). Each hugging face account has 10 "tokens" to use across all demos. Interacting with the demo will use one token. If you would like more tokens email [email protected].
Prototype 1: The NYC Decarbonization Data Chatbot
This demonstration provides a natural language bridge to NYC’s massive building datasets. It allows users to ask complex questions about building stock, energy conservation measures (ECMs), and LL97 penalties without writing code. This use case is intended to show the value that GenAI workflows can bring to analysis with large semi-structured datasets and highlights the value of these publicly available datasets.
A preview of the prototype in action can be found below. This was the demonstration presented to the working group. We have recently upgraded the underlying model to use stateful code execution. Rather than relying on static, predefined tools, the agent now dynamically writes and executes code tailored to each query, enabling near real-time answers while maintaining the same overall architecture. Check out the upgraded demo here.
The original model is still available here!
Prototype 2: Pre-Due Diligence & Risk Assessment Tool
This demonstration shows the application of LLMs as a "Pre-Due Diligence Screener" that automatically generates a preliminary Building Risk Report for any specified BBL in a few minutes. The tool acts as a "First Pass" for acquisitions or asset management, providing high level risk scoring by categorizing buildings into Low, Medium, or High Risk based on the available public data for that BBL, and a detailed report which covers everything it found about the building that could result in future risks. This assessment includes a summary of the buildings permit history, compliance history and complaint history along with a visual of the buildings submitted permits, violations and complaints.
A preview of the prototype in action can be found below. Try the model for yourself here!
Prototype 3: Peer Energy Use Evaluator
This demonstration utilizes the LL87 Audit dataset (which contains detailed mechanical system data and end-use breakdowns) to contextualize a building's performance. The tool automatically identifies a cohort of the 100 most similar "peer" buildings based on primary property type, vintage (year built), and square footage. It compares the target building’s energy end-use breakdown (heating, cooling, lighting) against the peer average. Then it cross-references the building’s permit, violation, and complaint history to identify targeted operational/mechanical deficiencies, such as recurring boiler repairs or elevator complaints, that correlate with high energy use. Using this information in combination with looking at what worked for similar peers in the cohort, the tool recommends Energy Conservation Measures (ECMs) with data-backed estimations of potential carbon and cost savings.
A preview of the prototype in action can be found below. Try the model for yourself here!
Prototype 4: Certificate of Occupancy Reader
This simple proof of concept takes any Certificate of Occupancy (CO) PDF, in any condition (scanned or digital) and extracts the table from the document converting it into an organized excel spreadsheet for the user.
A preview of the prototype in action can be found below. Try the model for yourself here!
Prototype 5: Drafter-Critic-Judge Implementation
This demonstration showcases a multi-agent architecture designed to improve reliability of these LLMs that was popularized by Andrej Karpathy, a founding member of OpenAI, who argued that agentic self-checking systems must exist as both the creator and evaluator of their own output. This system is often known as the Drafter-Critic-Judge (or LLM Council) framework where one model drafts a response, another model fact checks the response and a third model rewrites the response, taking in the critiques of the secondary model.
A preview of the prototype in action can be found below. Try the model for yourself here!
Prototype 6: Chat with Your Drawing
This demonstration presents an Image-Based Retrieval Augmented Generation (Image RAG) system designed to enable natural language interaction with a full construction drawing set. The system allows users to ask targeted questions about architectural, structural, and MEP content and receive answers grounded directly in the visual information contained in the drawings. In order to better interpret the drawing set, the model has been given the ability to identify relevant sheets and systematically narrow its focus to specific regions of those sheets.
A preview of the prototype as shown to the working group can be found below. The model has since been update to include Gemini's agentic vision! The updated model can be found here
The original model can be found here!
Prototype 7: Code Compliance Checker
This demonstration provides a proof of concept workflow that can verify a user's code related query against legal requirements for a drawing set. It implements a multi-LLM workflow to improve reliability by delegating distinct subtasks to specialist sub-agents. A central planner agent coordinates three expert agents: a Text Expert for rapid OCR-based reasoning tasks, a Page Expert for high-resolution visual verification using the Image RAG methodology described in the prior demo, and a Code Expert for structured retrieval and synthesis of NYC DOB code requirements. This demo was designed as a transparent, user-facing workflow. The interface exposes the interaction between agents as it occurs so working group members can observe the workflow in action.
A preview of the prototype as shown to the working group can be seen below. The model has since been updated to include Gemini's agentic vision and concurrent code/image retrieval. The updated model can be used here!
And the originial model can be found here!