Trustworthy AI in Real Estate

GenAI models have revolutionized the programmatic generation of text, following prompts in conversational interfaces.?

The Large Language Models (LLMs) that power these GenAI APIs and applications are amazing at the one task of predicting the next token. A token can be a word as a sentence gets built or even a set of characters as a word creates or auto-completes.??

The LLMs are machine learning models trained and fine-tuned on tons of publicly accessible data, specialized in predicting what should be written next, given all the texts fed into LLMs to train it.?

GenAI errors

In these LLMs, truth is not a variable; the words just need to make sense. When these models lie, this error is called “hallucinations,” in the context of an enterprise, these hallucinations can lead to flawed analysis and wrong decisions.?

The main reasons the LLM hallucinate are: (i) the training data did not include the necessary information to respond to your prompt correctly, (ii) the action requested from the LLMs is outside of its capabilities, and the model was not trained to decline to answer or, (iii) the statistical nature of these models renders a nonsensical answer from time to time.

Taming hallucinations on UDP

For an enterprise GenAi model to be trustworthy, it needs (i) access to your data, (ii) an ontological representation of the data that gives it meaning, and (iii) it needs to coordinate with other targeted programmatic features to get the job done.?

That’s a lot, so let’s unpack one argument at a time.

?First, the training data used to create the LLM does not include your data. Your business data is private and secured so that no one person or program can access it without permission. So, whatever model of the world the LLM has, it does not include your finances, operations, marketing, asset management, etc.?

That’s when UDP comes in: having all your business data in one place can securely coordinate LLMs alongside your data so that answers are pertinent.?

Even with your data, the LLM still needs to become an expert in real estate. Luckily, UDP is a real estate expert.

For UDP to get data from all your PMSs, single point solution systems, and benchmark market data, UDP has an ontological representation of your business: it understands what a lease is and how it relates to make-ready, occupancy, MTM rent, and all other relevant concepts. All your business data maps into UDP and makes sense inside of it.?

This ontology is provided to the LLM to fine-tune and fully train it to understand the concepts and relationships in our industry.?

Finally, the LLM may understand conceptually what a concession is but may need to learn how to calculate it, to which months to assign it, how to account for it on the effective rent metrics, etc. UDP has various programs that do such calculations and an orchestration layer between the LLM and the output that routes specific queries to the appropriate functions so the LLM can respond truthfully.?

In many operations, like creating new metrics and dashboards, UDP takes one step further, providing a human-assisted path and introducing intermediate steps and validations to ensure the output is trustworthy and can be shared inside and outside your organization.?

With this setup, the LLM writes text that makes sense, is easy to understand, and can translate complex graphs into simple words or turn simple words into complex metrics. General-purpose LLMs do not carry industry-specific knowledge and are blind to the specifics of your business. With UDP, owners and operators can understand the past with our BI tools, explain the present in simple terms, and see around the corner for what’s ahead.?