Just a few years ago, getting data reliably out of a PDF meant six months of OCR gymnastics, a tangle of regex, and a healthy dose of prayer. In spite of a decade plus of advances in deep learning and numerous document processing vendors emerging, document processing still had not *really* been solved beyond the simplest of use cases. Yet, so much of the world still runs on documents like PDFs, and documents are arguably still the standard communication protocol in industries like finance, supply chain, healthcare, and more.
Luckily, advances in large, pre-trained models seem like they have finally made a dent in this space. We now live in a world where you can just call an OpenAI API and do very complex classifications and extractions on fairly complex documents. Indeed - the improvements here have been so significant that many now consider document processing to be a commodity.
Yet, upon closer inspection, things aren’t actually so simple. If all you’re building is document search or some kind of RAG-based document system - you probably don’t need much more than what a foundation model can provide. These are use cases where 80-90% accuracy tends to be fine. However, this level of accuracy is insufficient for many of the most valuable document use cases.
Imagine you are uploading a paystub to a fintech service that is using the extracted data to approve or reject you for a loan - accuracy and reliability really matter in this case. The cost of a mistake is high, but the value of being able to automate this sort of workflow is immense as opposed to having to wait for human review and asking the user to wait 24-48 hours to get an answer back. LLMs can get you started here, but they won’t get you to the reliability you need on their own.
In other words, while transformer models have greatly raised the floor of what is possible in document processing, building production-grade document workflows still requires immense degrees of investment into “document processing infrastructure”.
If you speak to companies with these sorts of mission critical, in-product document use cases, you end up hearing a consistent story. They have assigned a team of 5-10 engineers to the problem for 1+ year, and have built a huge amount of tooling around the VLMs and OCR models including:
1. Annotation tooling for their internal subject matter experts
2. Human-in-the-loop workflows to handle weird document edge cases
3. Reinforcement learning & fine tuning workflows to learn and adapt over time from user behavior
4. Evaluation workflows to ensure they are hitting 99%+ accuracy
5. Workflow orchestration to chain together a complex directed-acyclic-graph of different models, combining optical character recognition, document splitting, document extraction, document classification, and more
6. Logic to handle all the myriad edge cases in documents - like handwriting, signatures, strike throughs, and sprawling tables
Rather than solve the problem, foundation models actually exposed the problem. They gave teams a “taste” of programming documents like APIs, until they ultimately realized that the model layer is just the tip of the iceberg, and that high quality document processing is actually a systems engineering problem. Indeed, we found that for many companies - including some in our own portfolio - document processing was the single largest bottleneck for their product roadmap and revenue growth targets.
And so today, we’re excited to announce our Series A investment into Extend, the company that we think is best suited to solve this problem. Extend is a developer platform for building mission critical document processing workflows. Like a “Stripe for documents” - their product exposes a foundational set of document processing primitives that a software team can stitch together to build extremely high accuracy and reliable document processing workflows.
Extend is built on top of two critical strategic decisions that are essential to scaling a startup in document processing.
First, Extend targets sophisticated software teams building document-based workflows - not back-office labor automation or RPA. Think bill pay for Brex or EHR data ingestion for Flatiron Health. These sorts of workflows are extremely high value if you can get them right, but they are also the hardest to get right.
Second, Extend does not attempt to be a “black box solution” or constrain software engineers in any way. Indeed, a lot of the value that Extend offers is allowing software engineers to try, test, and evaluate different AI models or processing strategies. Extend accelerates and enhances your in-house build rather than constricting it, which is an absolute must have if you are serving the most sophisticated companies and teams.
The feedback we have gotten on Extend is some of the best we have ever received on a Series A company. Quite a few customers we spoke to said that they viewed Extend as an extreme competitive advantage. “The fact that these guys are not Series C or Series D at this point is baffling to me” was a representative piece of feedback. The best infrastructure companies don’t just solve an engineering problem but enable their customers to deliver a novel product experience they could not have otherwise built, and Extend fits this to a tee.
Remarkably, Extend has achieved all of this - selling to customers like Zillow, Flatiron Health, Brex, Opendoor, Square, and more - with a team that was ~5 people until recently and less than $2M raised. Indeed, Extend actually had more ARR than funding raised when we met them at the Series A!
I have had the pleasure of knowing Kushal, Extend’s CEO, for a few years and it has been remarkable to not only see the intensity and urgency with which he has built the company, but also how effectively he has navigated the idea maze. There are a lot of ways to go after document processing, and Extend managed to narrow in on exactly the right approach, abstraction, and customer segment after a few years of iterating.
The rest of the Extend team is equally remarkable. As a few fun examples - Ishaan briefly held a world record in blindfolded Rubik’s cube, and is a top competitive programmer in spite of being a self-taught SWE. Gus created one of the world’s largest benchmark sets of tabular data and was one of the founding engineers for AWS sagemaker. Eli, Extend’s CTO, is not only an incredible technical leader, but exemplifies what it means to care deeply about every single detail of the company - from product to the office layout to marketing.
This investment reflects our continued focus on core software infrastructure as well as AI . More importantly, it reflects our desire to partner with the most exceptional teams we meet. We couldn’t be more excited to partner with Extend and hopefully, craft a world where programming documents is no different than calling an API.
