AI & GPTs in Deal Review & Underwriting

[Iliya Muzychuk]

Hi All, 

Wanted to share this experiment I ran on training my own GPT to review deals easily from screenshots. All I do is drop a listing into it and it can analyze costs, financing, comps, and so forth as I train the model. This is one of the new features of GPT4. I've also asked it to create some simple renderings of those deals for reference only. 

I want to know where further to train the model and how further of a use it could be for deals, whether marketing, analysis, financing, design and so forth since for a penny you basically can a junior employee that can copy your work methodologies. 

Also would be happy to connect with anyone who is interested in GPTs, LLMs and Tech in general in our industry. 

Thank and have a great day,

[Iliya Muzychuk]

Quote from @Jonah Kaplan:

Quote from @Iliya Muzychuk:

Quote from @Jonah Kaplan:

This is really awesome, @Iliya! Using GPTs is something I've been experimenting with as well! I'm a software engineer and lead an AI company focused on real estate. When I've tried to use AI for reviewing deals I've found the big blocker is the quality/formatting of data going in (IE: computer has trouble reading the pdf or XLSX). 

It looks like you circumvented that by having it browse a web page, is that right? Also you mentioned you did some training of your own GPT. What resources did you use for the training? 

I used openAI's custom GPT builder, and uploaded engineering documents for it to review. It used Vision to read through some of them, some successfully and some less. But these days I've navigated more into MidJourney to do quick feasibility studies and renderings for realtors and investors. I've created a combo workflow of generative design with generative AI for these type of studies. 

What type of company do you have? Would love to learn more how you use AI in real esate! 

Oh very cool. So maybe more on the design or development side? 

We primarily work with brokerage teams. We've built integrations across dropbox, box, Gsuite, Outlook, etc. to bring all of that data in and give teams the ability to search and ask questions over their doc. The big piece is connecting work together (ie: this property in salesforce relates to these docs across dropbox). We've also just started playing with fine tuning an LLM to be real estate specific - so that it can better get in to analysis. This is why your post really caught my eye.

Is this an area you're going to continue exploring? Curious to know your take on it!  

I think LLMs in real estate are promising just due to the shear amount of information we share in the creation of permit/construction documents. From an architecture stand point, we use tremendous amount of textual data such as zoning codes, contracts, building codes, proposals, specifications, communication logs, meeting notes and more. You'd think it is drawings that take the space, but most of the work is administrative (70% or so). For example my previous firm had about 8TB of data for just current projects, that gives you a sense of scale. 

We spend so much time tracking back the data and decision reasoning chasing like rabits through folders to find the reasoning of changes and omissions. Any minor change could be a reason for legal action in the future so we have to keep everything on records for years. 

LLMs definitely could be super handy, and also help us increase the quality of the projects and consturction documents we produce. Now when everything is moving to BIM they probably can even query the models and give us better insights as well. 

I am definitely exploring where this technology can be handy, started my own path and seeing the benefits of using generative image creations, language models, and even co pilot to write simple code for my design models. There is a field called generative design that combined with generative AI create powerful workflows to deliver design documentation quickly and I am just wandering curiously in that domain at the moment.