
If you prefer reading instead: extend.ai/resources/opendoor…

Document infrastructure for agents. Trusted by startups and F500s to turn their most important documents into data.

If you prefer reading instead: extend.ai/resources/opendoor…

This week’s customer highlight is @Opendoor 🏡 Back in July, we sat down with Yan Lhert from their title and escrow team. Each transaction involves 200 to 300 documents, and his team built an AI QC system on Extend to review them, reaching production-ready performance in days. Watch how they built it below!

look, we get it — building good benchmarks is extremely hard...but that’s exactly why benchmark slop needs to be called out we notice this frequently from the company behind ParseBench, and ignored it the first few times, but don’t think its fair to customers, our competitors, or the broader community who look at benchmarks to make informed decisions selectively choosing datasets, gaming scoring methodology, or hastily throwing stuff together that’s riddled with errors just so you can appear #1 as a marketing tactic is disingenuous. While we tell users to eval on their own data, not everyone does and benchmark posts (even if flawed) shape opinions. parsebench heavily favors llamaindex, who created it, and like their other benchmarks has a ton of issues with it: - penalizes small formatting errors that users don’t actually care about - its overfit to their models, and introducing small amounts of variance and noise will dramatically change the results Extend has its own benchmark, RealDocBench, that we spent a lot of time and effort into creating: - we built it around agent Q&A, because that’s what users actually care about. nobody’s agent breaks because a heading was missing an underline or had a few extra spaces. instead, we ask whether an agent can answer questions correctly on parsed output from production-grade healthcare, real estate, finserv, etc documents - we did our best to make it fair. its why our competitors also do well on it (e.g. reducto’s r-1 performs well matching our performance parser, r1-mini matching light, etc) this is a massive market that all of us can compete in. long term, helping users make informed decisions is best for all of us it’s not worth sacrificing your credibility for short term marketing tactics


multiple prospects told us this month that other providers quoted them $25k+ just to turn on zero data retention ZDR costs a provider almost nothing. it’s the absence of retention, not a feature. it just happens to be a hard requirement in regulated industries, which makes it a perfect toll booth on Extend there’s no ZDR tax. it’s included on every paid plan (even pay-as-you-go) with no annual contract or platform fee, just ask us to turn it on our customers trust us with the docs their business runs on: mortgage apps, healthcare records, bank statements. that trust shouldn’t come with an add-on fee your data is yours
View quoted post ↗
multiple prospects told us this month that other providers quoted them $25k+ just to turn on zero data retention ZDR costs a provider almost nothing. it’s the absence of retention, not a feature. it just happens to be a hard requirement in regulated industries, which makes it a perfect toll booth on Extend there’s no ZDR tax. it’s included on every paid plan (even pay-as-you-go) with no annual contract or platform fee, just ask us to turn it on our customers trust us with the docs their business runs on: mortgage apps, healthcare records, bank statements. that trust shouldn’t come with an add-on fee your data is yours

pov: me when free sauce gets posted.
the most used internal tool at @ExtendHQ started as a voice note on a facebook marketplace peloton here's how growth lead @the_doc_doctor built our GTM brain with claude code!


all of you yelling about company GTM brains, read this to see how we do it at Extend @the_doc_doctor goat
the most used internal tool at @ExtendHQ started as a voice note on a facebook marketplace peloton here's how growth lead @the_doc_doctor built our GTM brain with claude code!


Read how @the_doc_doctor built our internal GTM brain below!
the most used internal tool at @ExtendHQ started as a voice note on a facebook marketplace peloton here's how growth lead @the_doc_doctor built our GTM brain with claude code!


the most used internal tool at @ExtendHQ started as a voice note on a facebook marketplace peloton here's how growth lead @the_doc_doctor built our GTM brain with claude code!



One of OpenAI’s latest models looked at this invoice and called it legitimate with 95% confidence. The math was correct. Except the fraud wasn’t in the math. Someone altered the PDF and inflated the PO balance by $30k. Verify caught it by checking the file structure, not just by looking at what's on the page. If you’re building agents for financial services, lending, or insurance, you can’t blindly trust an approval like this. Your agent can reason perfectly over a forged doc and still approve it, and one bad approval can cost tens of thousands of dollars and your customers’ trust.

If you want to build a Claude Sheets-like experience (+ fast parsing & accurate charts for xlsx, xls...) into your own app, check out github.com/extend-hq/react-x… v0.21.0 (MIT) Took some inspo from Claude's interaction pattern here but you can customize the menu & callbacks
We just added Whiteboards and Sheets to Claude, and Dashboards is now Claude Data. Work with your team and Claude to: • Sketch a plan and brainstorm live • Ask for a formula in plain words • Turn the numbers into a story Sketch, sum, show it in Claude

want to embed a Claude Sheets-like experience in your own product? Extend UI has a drop-in spreadsheet UI with fast parsing, accurate charts, customizable interactions open source + MIT licensed as always
If you want to build a Claude Sheets-like experience (+ fast parsing & accurate charts for xlsx, xls...) into your own app, check out github.com/extend-hq/react-x… v0.21.0 (MIT) Took some inspo from Claude's interaction pattern here but you can customize the menu & callbacks

i recreated Papers Please, but you're deciding which YC startups to fund your job is to catch founders deceiving you with their financials and documents, and decide who's worth investing in it's called Term Sheet Please: termsheetplease.com/

table parsing is like reading a spreadsheet out loud to someone else who is typing it out every cell, every border, one word at a time @joon_kim96 worked his inference magic (or spec decoding if you want to be specific about it) and cut table parsing by 4x see how he did it
we optimized inference for our table model and made it 4x faster on a single H100, median latency on our Light table model dropped from 3.35s to 0.76s our ML engineer @joon_kim96 wrote up our learnings and how we ultimately ended up using DFlash:
