Reducto: Making Human Data LLM-Ready With State-of-the-Art Accuracy

1. 0:00 – 1:17

   Data-driven AI for large enterprises

   1. DHDiana Hu0:00

      I'm excited today to welcome the founders of Reducto. They went through the YC batch back in winter '24. So welcome, Aadit and Ronak. Tell us a bit about what Reducto is.

   2. SPSpeaker0:15

      Sure. Um, so at Reducto, we help people take their really, really complicated documents, insurance claims and health records and financial statements, and we turn that into clean structured data that they can use for any sort of use case, um, but primarily LLM-based use cases like RAG and summarization.

   3. DHDiana Hu0:32

      And to give a bit of a context for the audience, you guys made a bunch of progress and actually have a lot of, uh, large enterprises using it.

   4. SPSpeaker0:39

      Yeah.

   5. SPSpeaker0:39

      Sure. Um-

   6. DHDiana Hu0:40

      Some of them that you can maybe name drop.

   7. SPSpeaker0:42

      Yeah. Um, so honestly, we've been fortunate to get to work with some really cool teams in the space, uh, companies like Vanta, um, and many more. Some of them we can't name, but it goes all the way up to enterprises in the trillion-dollar market caps, um, where I think this is a really core and fundamental bottleneck for people building AI applications today. Um, so even though it's been a short period of a year, uh, things have just progressed really quickly on that front.

   8. DHDiana Hu1:09

      What's been impressive is you guys have been one of the AI companies that have gone through YC that have seen that really quick growth just within a year.

2. 1:17 – 3:04

   Document management

   1. DHDiana Hu1:17

      How did you come up with the idea? I know there was a bit of a wandering in the wilderness for a little bit.

   2. SPSpeaker1:22

      Yeah.

   3. DHDiana Hu1:22

      You guys had applied with YC with a very different idea that was a open source tool, and then eventually landed on Reducto.

   4. SPSpeaker1:30

      Yeah. Um, so it started, like you mentioned, with, uh, something very different. Um, Ronak had built long-term memory for LLMs, and he was actually, I think, one of the first people to do that ever. Um, but at that point, it was sort of this, like, really cool thing to see, but not something that anybody was trying to implement in their applications. Um, but over the course of that, one of the features that's we were sort of working on is people would mention, "Hey, you're managing our user chat history. Um, can you manage the files that users are uploading too?" And we tried just using off-the-shelf solutions for that. Nothing really worked very well. And so what Reducto is today wasn't supposed to be a pivot. Uh, we had this, like, almost embarrassing version of a segmentation model, like just thrown into a Streamlit app, thrown on BookFace as like a-

   5. SPSpeaker2:21

      It was literally a weekend project, like-

   6. SPSpeaker2:23

      Yeah

   7. SPSpeaker2:23

      ... slopped together with a bunch of heuristics and, and things like that.

   8. SPSpeaker2:26

      Yeah. Um, and that we just put up as like a blog post of how to segment your documents effectively. Um, and again, it wasn't supposed to be a pivot, but surprisingly, that feature that was supposed to be a marketing stunt immediately had people saying like, "Hey, this is better than what I'm getting from, you know, Textract. Uh, can you make this an API? Is this something that I can just pay for on Stripe?" And I, I think over the course of talking to those teams, we really started to see that teams that have no sort of reason to be PDF processing companies were spending a lot of their time just on the sort of data ingestion piece that to them is just a bottleneck from doing the things that they really care about.

3. 3:04 – 3:59

   Simplify PDF processing for companies

   1. SPSpeaker3:04

      So we sort of abstracted the problem of like how can we be their ingestion team for them.

   2. SPSpeaker3:09

      Yeah. None of these like AI application layer companies like want to be PDF processors. It's just not something that's like exciting to them, and that's something that we like are excited about just abstracting away the problem. We really say like we try to be the ingestion team for the companies that we work with, um, and try to build our company in a way that works well for them.

   3. DHDiana Hu3:28

      I think what's fascinating about this idea-

   4. SPSpeaker3:30

      Yeah

   5. DHDiana Hu3:30

      ... I think you both are very, very hardcore engineers, and you could have been in that camp of like, oh, processing PDF is like a simple problem. But the funny thing is that you were surrounded by a community of founders building AI products, and they all still were struggling with it. There wasn't nothing around it that worked, but nobody wanted to do it because it's kinda like boring stuff.

   6. SPSpeaker3:53

      Yeah. Honestly, like we were surprised too, um, that the problem wasn't kind of solved.

4. 3:59 – 5:02

   Aha moment for PDF extraction, interesting approach

   1. SPSpeaker3:59

      Um, when we were first looking into the problem, we tried all of the existing solutions on the market, and I was expecting something to just work, but it didn't, and that was kind of the aha moment of like, hey, we should maybe invest in, in this ourselves and try to solve this problem for other people.

   2. DHDiana Hu4:13

      Kinda reminds me a lot of, uh, the famous PG essay, the schlep blindness.

   3. SPSpeaker4:18

      Yeah.

   4. SPSpeaker4:18

      Yeah.

   5. DHDiana Hu4:19

      It's one of those things that all developers and AI engineers know about this, that PDF aren't... extraction aren't 100%, but we don't wanna work on it. It's kind of boring. Same thing with like the example that PG points in his essay is, uh, Stripe. It's like, oh, people knew back then that payments processing was kinda annoying, and the API was kinda eh, but nobody really wanted to work on it, and you guys jumped on it, and it was good.

   6. SPSpeaker4:43

      And-

   7. SPSpeaker4:43

      Yeah

   8. SPSpeaker4:43

      ... I think part of the approach that we took was we, we, we made it a more interesting problem because we took a different approach than what a lot of other folks had been doing in the sense that we turned PDF processing, which is usually just writing a bunch of rules for how to process this type of file type and this other type of file type into a computer vision

5. 5:02 – 6:56

   NLP-based PDF extraction for enterprise apps

   1. SPSpeaker5:02

      problem. So the insight we had is like we are gonna parse and understand these documents the way humans do, and that's really what my background was in, was like machine learning research, computer vision into autonomous driving and things like that. Um, so I was really excited at the idea of applying a lot of the techniques from the computer vision space like to, um, processing documents kind of the same way humans do.

   2. SPSpeaker5:23

      Yeah.

   3. SPSpeaker5:23

      And that was the start.

   4. SPSpeaker5:25

      Yeah. I think if you look at it as this sort of first principles problem, um, there's so much information that's to us is very intuitive, right? Like every gap between two paragraphs is me telling you, "Hey, this is like a new piece of semantic information," um, or a tab is like me telling you, "Hey, this is a nested hierarchy." And when you sort of boil it from that lens of like how do you try to read not just a specific type of invoice, but anything that could be on the page, um, it's a really hard thing to do, uh, but it's a fun thing to sort of work across that long tail.

   5. DHDiana Hu5:58

      The cool thing that I remember you guys figure out during the batch was that by becoming the best PDF extraction tool-A lot of AI enterprise applications just got better. You guys became this foundational piece of the new AI application layer-

   6. SPSpeaker6:15

      Yeah

   7. DHDiana Hu6:15

      ... for enterprises, and you saw a lot of your customers succeeding with you. Maybe can you tell us about the gains they saw by using you?

   8. SPSpeaker6:22

      Yeah. This is actually my, like, favorite category of customer Slack message. Um, people will will often see end LLM accuracy improvements just from swapping the ingestion provider, uh, 'cause the data that you provide in is fundamental to what's being reasoned on. Um, the, like, extreme end of the cases, we've had customers report plus thirty percent, uh, sometimes even more if they're re- dealing with really challenging documents. Um, but for almost everyone, there's the benefits that you get from better accuracy. We take a lot of the work of post-processing out of the plate for them too, like chunking and all that kind of stuff.

6. 6:56 – 8:10

   Great data, exciting use cases

   1. SPSpeaker6:56

      Um, and then the question basically becomes, okay, you have great data. What do you do from there?

   2. SPSpeaker7:01

      My favorite category of, like, improvements or messages from customers is, like, this use case wasn't even possible before because users were uploading scanned documents that didn't have metadata, or we needed to process, I don't know, hundreds of pages in, like, a few seconds or something like that, that, that just no approach was capable of doing before. And so all of a sudden they're able to implement new features into their applications and products that just weren't possible before. That's the most exciting thing, to be able to, like, enable those, like, new use cases for customers.

   3. DHDiana Hu7:29

      And you guys shipped a lot of this very quickly. I do remember one of the examples I saw from you guys. It was this, uh, when people buy homes, there's this giant-

   4. SPSpeaker7:39

      Yeah, the questionnaire

   5. DHDiana Hu7:40

      ... the questionnaire that people, real estate agents fi- file by hand, and it's like garble handwriting. And you guys were able to extract it a hundred percent correct.

   6. SPSpeaker7:49

      Yeah. Um, I do think that this is something that's only possible today. Um, we, we get the benefits of not just being able to use traditional computer vision to get, like, bounding boxes and deterministic outputs, but also to capture that long tail where if you have a snippet of handwriting, VLMs are so much better than traditional OCR was there.

7. 8:10 – 8:48

   Best places for customer approaches

   1. SPSpeaker8:10

      Um, and going back to the sort of analogy of we are the ingestion team, I think it's on us to sort of find the places where we can get the best out of each approach, uh, such that our customers don't have to think about whether their questionnaire has, you know, check boxes or handwriting or any of that.

   2. SPSpeaker8:25

      And it's even the case where customers will come to us with like a very specific kind of thing, where maybe a cell within a table has a green highlighter circling a-

   3. SPSpeaker8:34

      Yeah

   4. SPSpeaker8:34

      ... a number. Um, and so that's the kind of thing that we're really excited to be able to, um, handle, where users can even just instruct or guide our models, um, by giving them like plain text instructions to make changes, things like this.

   5. DHDiana Hu8:47

      Mm-hmm.

   6. SPSpeaker8:47

      Yeah.

8. 8:48 – 11:21

   Closing a Fortune 25 deal in just two months

   1. DHDiana Hu8:48

      So tell us a bit about how within just a couple months of graduating from the YC batch last year, you guys were one of the few companies that quickly closed a Fortune 25 company to become a full-on enterprise customer.

   2. SPSpeaker9:08

      Yeah.

   3. DHDiana Hu9:08

      Not just going through the whole pilot procurement security, but actually fully signed the contract.

   4. SPSpeaker9:13

      Yeah.

   5. DHDiana Hu9:13

      That's the fastest I ever seen. How was that journey?

   6. SPSpeaker9:17

      Yeah. Um, it was intense. Um, to us it felt very long, but I hear in retrospect that it's somewhat fast. Um, so that actually started during the YC batch. We did our launch YC. Um, to us, I remember before we did our launch, we were very big on at least something has to be substantially better as a result of what we're releasing, and we scoped that as we should have a really, really good layout model that will decompose your document for you. Um, and so they tried a document that I think was failing in their internal pipeline, and that worked, and so they took that initial demo call. Um, we were ecstatic. Uh, I think we sent you a screenshot of like the form submission.

   7. DHDiana Hu9:55

      [laughs]

   8. SPSpeaker9:55

      I guess there was this, like, constant learning because neither of us are salespeople. It's not our background. Um, but really it just became this question of, like, can we show to their team that we will help their team do something else faster? And, uh, going back to that same point, like they had their own internal document processing team. Um, that was our real competitor. Like they tried off-the-shelf tools, but we were competing with people that had access to their distribution of data. Um, they knew what they were evaluating against, and this was all stuff that they wouldn't reveal to us.

   9. SPSpeaker10:25

      They were fine tuning on the examples they were, like, drafting on basically, right? And we still had to-

   10. SPSpeaker10:29

       Which is a hard bar to, to try to capture.

   11. SPSpeaker10:32

       Yeah.

   12. SPSpeaker10:33

       Um, and so yeah, it was this five-month ordeal where at first it would be one meeting and then they'd say, "Okay, like, maybe let's also meet this other person. He wasn't in the room today." Um, and it kept going and then going until at some point we spent eight hours in their office for the full day with like fourteen people from their team, um, of them just grilling us, including the team that we were competing against, who were, like, kind of trying to pick at how we got to where we were. Um, but fortunately it ended up working out. So we've been working with them for, uh, I guess about half a year now-

   13. SPSpeaker11:03

       Mm-hmm

   14. SPSpeaker11:03

       ... uh, and continuing on.

   15. DHDiana Hu11:04

       That's very cool, and I think part of it is, I think, Ronak, you were able to get Reducto to build the best model ever for document extraction, and that was quite a journey, and I think you guys released a data set and a benchmark and tried to set the standard.

9. 11:21 – 13:19

   Data-driven AI for high-quality documents

   1. DHDiana Hu11:21

      Tell us about how you guys got, got to basically SOTA.

   2. SPSpeaker11:24

      Yeah. Um, it's a hard problem, especially in this space. We were expecting there to be a lot of really good quality data sets for us to, to measure up against. Um, but we often saw that in the space of documents, like people just weren't often willing to put in the, like, effort that was required to generate the really high-quality data. So part of that process is in-house, we, um, reached out to, like, teams of PhDs, and we actually, like, built really, really extensive, like, data pipelines and like kind of a data engine in-house for us to be able to sample some of the most, like, diverse document data that we could find. That basically allowed us to start iterating a lot faster, um, on our end, um, where we could, um, we could test on any given case, like, how are we doing, um, how can we compare against all of these other methods? And part of it, I think, is alsoUs coming into the equation without, um, kind of maybe a lot of the baggage of like knowing how people were parsing these documents before, um, using a bunch of heuristics and manually parsing these like PDF document standards. Um, coming it-- into it with like a new and fresh perspective of like, "Hey, we're vision first. We're gonna read the document like a human does." That, I think, unlocked a lot of like maybe novel approaches, um, that folks might not have considered before.

   3. DHDiana Hu12:42

      And you guys have expanded beyond PDFs, right?

   4. SPSpeaker12:45

      Yeah.

   5. SPSpeaker12:45

      We do spreadsheets, images, documents, slides at this point.

   6. DHDiana Hu12:49

      That's very cool. So you've become the, uh, the standard for a lot of, uh, now enterprises that need to process all these tools that are adopting AI is-

   7. SPSpeaker12:56

      Yeah

   8. DHDiana Hu12:57

      ... the tool to use.

   9. SPSpeaker12:58

      Yeah. Um, people don't wanna fork their pipelines. Um, in the beginning, we really wanted to constrain our focus because we're a small team, and we need to maintain this accuracy bar that we set for ourselves. Um, but we would just constantly get requests of like, "Please build a spreadsheet endpoint."

   10. DHDiana Hu13:14

       [laughs]

   11. SPSpeaker13:14

       Like, I am building this whole separate pipeline. And so we started giving those options

10. 13:19 – 15:18

    Reductos AI-focused infrastructure attracts top companies

    1. SPSpeaker13:19

       too as well.

    2. DHDiana Hu13:19

       So one of the things that's really cool to see is that you guys are becoming a key piece of infrastructure for AI applications and a lot of AI agents, and a lot of the best YC companies are choosing you guys. Tell us about why is that, and if someone is building an AI tool or application, why should they use you?

    3. SPSpeaker13:42

       The, like, big unlock here is the only reason people were building this in-house is because it was stopping them from getting their products to the quality bar that they were looking for. And what we are trying to do is make it so much faster for them to, you know, work with the newest models, to focus on, you know, improving the post-processing and reasoning after that. Um, so for honestly any team, whether it's a startup, scale-up, or enterprise, um, I really think that Reducto is sort of like core infra that hopefully they see as a tool for them to leverage and then build off of into anything that they're looking for from there.

    4. SPSpeaker14:19

       Yeah.

    5. DHDiana Hu14:21

       And as part of your fundraising announcement for your Series A, congratulations.

    6. SPSpeaker14:25

       Thank you.

    7. DHDiana Hu14:25

       You guys are growing the team a lot. If someone wants to work for you, what, what, what kinds of engineers are you looking for?

    8. SPSpeaker14:31

       I think we're hiring for basically across all of our engineering roles at the moment, product, machine learning engineer. Um, the type of people we tend to look for are really scrappy, have been like startup, like, founders or, uh, engineers in the past. Um, and the thing that we really love to index on is just like really caring about the details. Um, w-we've hired some folks and, like, our first machine learning researcher, um, first day he started, like, training a model, and I just saw him sitting at the desk and, and like clicking through like thousands of pages of, of documents. And I think part of... That like surprised me at first, but that's kind of the quality that you need to have when you're really building something cutting edge. Like

11. 15:18 – 15:34

    Quality of data, results, support

    1. SPSpeaker15:19

       Reducto, for us, it's extremely true. For our customers, it's also true. Like the quality of your data is the quality of your end outputs and results.

    2. DHDiana Hu15:27

       All right, guys. Thank you so much for coming and chatting with us. And again, congratulations on your fundraising announcement for your Series A.

    3. SPSpeaker15:33

       Thanks, Diana.

Episode duration: 15:34