There were moments when it felt like Extend was part of our team. You would forget that they were not even working here.
Yan Lhert, Staff Engineer, Opendoor
Opendoor exists to tilt the world in favor of homeowners and those hoping to become homeowners. They've served 300k+ homeowners by making homeownership simpler, faster, and fairer. An average home transaction will involve 200–300 documents, each of which can be up to 200+ pages.
Yan Lhert is a Staff Engineer at Opendoor, where he works on the title and escrow team. We spoke with Lhert about why Opendoor stopped trying to build document processing in-house, how the team got to near-production in days, and what it looks like to run 50,000 documents a month through Extend.
Where do documents come into the picture at Opendoor?
Lhert: Buying and selling homes is extremely document-heavy. A home transaction will involve somewhere between 200, 300 documents. Most of the context is locked inside of PDFs. PDFs are the API that run home sales.
What was document processing like at Opendoor before Extend?
Lhert: Opendoor historically has spent most of its document processing time on humans, manually typing out stuff that's in a PDF. It's mostly comparing documents against each other. Extremely cumbersome process, rife with errors.
What's at stake when document processing goes wrong?
Lhert: The stakes are extremely high. It could be as much as the full loss of a property or very serious litigation. The build-versus-buy equation is more biased towards building than ever, and we ran into a number of accuracy issues when we were trying to build this ourselves. Things like bounding boxes, eval sets, splitting things, tables that span multiple pages. Quickly we learned we need a better tool that can help us.
How did you evaluate the options? What did you benchmark against?
Lhert: We started out looking at some tools like Textract and moved on to some of the other startups in the space. Extend was by far the best in terms of all-around performance and easiest to set up, best team to work with. It was kind of a no-brainer.
Did you consider building it in-house?
Lhert: One of the other engineers on our team, Dharma, sat me down and said, "Hey, why don't we just build this ourselves?" It didn't take long before he came back around to me and said, "Yeah, we should just use Extend."
What differentiated Extend in your testing?
Lhert: The docs were incredibly good, and with llms.txt we were able to throw it into our favorite Claude Code session and get rolling right away to near production experience with the product within a matter of days.
Which use cases have you rolled out on Extend?
Lhert: Extend is now used in a number of production paths at Opendoor. AI HUD QC allows us in title and escrow to verify thousands of pieces of critical information. It's been a huge win for us.
Has Extend materially impacted Opendoor's business?
Lhert: At this point, in even just a few months' time, it has put hundreds of thousands of dollars into our bottom line. And it's prevented us from making mistakes that, if a human had been looking at those documents, it's very likely they would have made a mistake.
If Extend vanished tomorrow, what would be the impact?
Lhert: I'd probably have to quit Opendoor if Extend vanished tomorrow, because I've learned to rely on it for so many things and I absolutely love using it.
