Duxre, a platform for commercial real estate transactions, announced the launch of 'AI Brief'. This software converts unstructured sales and leasing data into a readable format for large language models (LLM). Founded in 2020 and based in Santa Monica, California, Duxre offers a subscription platform designed to optimise data processing for commercial real estate brokers, improving transaction volume and client engagement.
'AI Brief' converts unstructured files into Markdown files, which can be read more efficiently by LLMs such as ChatGPT and Claude, according to the company. This reduces file size from 50 megabytes to 30 to 40 kilobytes, accelerating processing and minimising errors.
Challenges in Data Processing and Duxre's Solution
Sean Fulp, founder and CEO of Duxre and Vice Chair at Colliers (the two companies are not affiliated), explained that the idea for 'AI Brief' stemmed from experiences marketing properties for sale and lease. He observed that typical deliverables, mostly PDF documents, are fed into ChatGPT or Claude for deal analysis. An offering memorandum, which can be 60 to 80 pages long and 15 to 20 megabytes in size, is therefore no longer first read by a human, for whom these documents were originally created.
Fulp pointed out that there is no sophisticated pipeline for the real estate industry that provides an evaluation system for these documents to ensure the accuracy of the extracted information. He observed that LLMs tend to embellish and 'hallucinate' documents, leading to problems in the process. LLMs could provide convincing but incorrect results to tenants, buyers, or brokers. This could even widen bid-ask spreads due to misinformation.
How 'AI Brief' Works
'AI Brief' operates in three stages. First, there is multi-source ingestion, where the brief is created from all information contained in the deal file, not just a single document. This includes offering memorandums or leasing brochures, Excel financial data, and structured listing data that the broker has already entered, such as address, property details, and further descriptions.
- —Synthesises and reconciles data from various sources.
- —Uses code and AI, with language processing handled by a model, while mathematical calculations and structure are managed by deterministic code to calculate figures directly from the source.
- —An evaluation layer assesses the output before submission to verify the accuracy and completeness of the extraction.
According to Fulp, the key difference is that a Markdown file is 30 or 40 kilobytes in size, compared to 50 megabytes. The file size is thus about 95 per cent smaller than an offering or a leasing brochure. A smaller file size allows for significantly faster loading speeds in ChatGPT, Claude, or Gemini and substantially reduces costs. The cost of processing a PDF can be 15,000 to 20,000 tokens, whereas a Markdown file only requires 1,000 or 2,000 tokens.
Additionally, 'Marketplace Duxre' on the Duxre website offers a free subscription for creating sales or leasing offerings. Brokers can publish a listing and generate a brief for it.
