đ§đ From Pivot to Fortune 10 Customer, Lessons in Founder-Led Sales with Reducto CEO Adit Abraham
Inside Reducto's 48-hour Series B, why founders should demo on the first sales call, and how to avoid pivot hell
Reducto went from YC to closing a $75M Series B in 18 months, burning only $1 million in capital along the way.
This latest episode of The Peel goes inside Reductoâs big early pivot to building the best product for processing documents, and how they avoided pivot hell to get there.
Adit Abraham, Co-founder and CEO of Reducto, shares how they landed an early Fortune 10 customer as a two person startup, everything he learned about sales growing Reducto to its first $5m in ARR before hiring any sales reps, and how a redeye flight and crazy hustle helped hire their first AI researcher.
Support this Episodeâs Sponsors
Numeral: The end-to-end platform for sales tax and compliance. Try it here: https://bit.ly/NumeralThePeel
Hanover Park: Modern, AI-native fund admin at https://www.hanoverpark.com/Turner
To inquire about sponsoring future episodes, click here.
đ Stream on Spotify and Apple
Timestamps to jump in:
3:35 Reading unstructured human data
10:44 Growing 5x in four moths
12:38 Insurance, healthcare, legal, logistics
19:13 Things LLMâs still struggle with
28:23 Starting Reducto from a blog post during YC
32:01 Landing a Fortune 10 customer
35:48 Limiting the product and growth early on
40:57 Getting a professor fired at MIT
43:50 How to avoid pivot hell
49:00 $108M from First Round, Benchmark, a16z
51:48 Chetan convincing them to raise a Series A
55:50 Raising a Series B in 48 hours
59:36 Redeye flight to hire the 1st AI researcher
1:05:42 Lessons hitting $5m ARR with founder-led sales
1:13:09 How to stay on top of changes in AI models
Referenced:
Find Adit on X / Twitter and LinkedIn.
Related Episodes
đ Stream on YouTube, Spotify, and Apple
Transcript
Find transcripts of all prior episodes here.
Turner Novak:
Adit, welcome to the show.
Adit Abraham:
Thanks for having me.
Turner Novak:
Yeah, I think this should be fun. Really quick for you who donât know, what is Reducto?
Adit Abraham:
So Reducto, it started as an API to be able to read any sort of documents for language model use cases. Today, more broadly, you can think of Reducto as a layer that helps you connect all sorts of unstructured human data, your PDFs, spreadsheets, whatever you might have, and interact with them for whatever you might need to do. That includes the parsing that we started with, but it also includes a lot of newer things like editing documents, getting structured insights, splitting and classifying, everything youâd need as a toolkit.
Turner Novak:
So what does that actually kind of look like practically, what kind of things could someone do with it?
Adit Abraham:
Every single person watching this thatâs probably at some point uploaded a PDF into something like ChatGPT. If youâre building those sorts of products, that context is really important and you need great accuracy, not just for the simple text PDFs, but things that maybe have tables and charts. The more complex cases because a lot of the more important points of human data is things like faxes, patient cell chart. So we will go through, we will parse all of that data out really accurately. Weâll structure it effectively for language model inference. Weâll do a lot of the post-processing that people would need so that when people pass into language model, theyâre getting the best answers that they can.
Turner Novak:
And why is that so hard? Because I would kind of think we have LLMs. ChatGPT came on a couple of years ago, this should be just fixed, right? This should have been solved a while back. Is it still such a big problem for people?
Adit Abraham:
I talk about this a lot. Where itâs not even like a LLM specific problem. People have been working on PDF processing longer than Iâve been alive. I know people that built the original drivers for printers to print PDFs back in the 1990s and there have been many generations of companies since. But the context of the problem is very different, which is historically, you would probably have a document processing vendor thatâs templated out one type of document, like just your bank statement for one type of bank or just your W2s. And these were brittle pipelines that youâd spend a lot of time maintaining and even deploying in the first place.
But todayâs context is completely on the other side, which is that people want to be able to any form of unstructured data. Weâre not far from a point where when you apply for a mortgage, thereâs going to be some loop that goes through all of your transactions and makes the judgments. When you go say hi to your doctor, thereâs going to be some summary of your medical records and when you need to read all of those, the hard thing is, at the end of the day, these documents are almost like a whiteboard that you could canvas any which way.
Often when we think of documents, we think of just that canonical paragraphs next to each other, but you could have things like patents, where you have two columns and you have line numbers and you start to see things like the line numbers mess up the data in the actual paragraphs of text. Youâll see things where... Youâll have financial tables and if there arenât clear grid lines, sometimes the language model will choose data from the wrong cell in the table.
All of these downstream problems thatâs... Obviously, since this is the first step of that pipeline, every mistake that you make at ingestion is going to compound throughout your pipeline. And so, one, this was an issue when people were building chatbots they needed to ask questions to. But two, now that people are trying to do end-to-end workflows that youâre trying to automate your entire invoice parsing, youâre trying to go from documents to a completed work product, you just need human level accuracy for that data and thatâs a much harder problem than the canonical traditional OCR.
Turner Novak:
Yeah, it kind of reminds me of when you just think about what computers and software do back in, I donât know, the â40s or the â50s with accounting, you just have to write all this stuff by hand or whatever, you do the books, credits and debits, et cetera. Maybe you had to erase things or whatever, tons of different pieces of paper and we basically just kind of took it and did the same thing on the computer. Itâs faster, but thereâs still a lot of, when youâre an accountant, youâll get a piece of paper and youâll type it into the computer and itâs still a lot of this kind manual human labor and itâs interesting.
I mean, if you just think about whatâs AI going to do, itâs just going to automate a lot of this manual human labor that doesnât really necessarily require a lot of skills or decision-making. Itâs just kind of a yes or no or a route tree type of thing. But to your point, really hard. So itâs kind of interesting when you just think about your infrastructure of doing this harder manual labor that weâve done for, I donât know, all of human history, basically.
Adit Abraham:
One, we see it across literally everything. As somebody that has been in tech for my entire career, I think I almost didnât appreciate how deep the problem goes, but when a large insurance company receives a claim submission, they receive just a folder with a bunch of different files when itâs all the supporting evidence. And they donât even know whatâs in that folder. Itâs not well tagged and itâs listed as, hereâs the data that youâll find in this folder or something else.
Turner Novak:
Itâs literally like a claim holder or a policy holder thatâs like, âOkay, hereâs a bunch of pictures, a video, my title for my mortgage or the deed for the car or whatever the claimâs for.â
Adit Abraham:
Just all over the place, yeah. And so, that was something that historically, you would need to get humans to look at. And Iâve heard crazy stories here where I was meeting with the CIO for a really large insurance company, like multiple billions in revenue, and he was talking about how for every 10 large claim submissions that they receive, they work with really large claims, they only have enough humans to look at three out of those 10. And of those three, only one or two will actually be applicable to what they want to cover. And so, thereâs this other 70% where it might be relevant, it might not, they just donât know because itâs been a human process up until now. And thatâs the kind of thing where even from the beginning, you could see the promise of language models in that space. From the early days, there was the scope of what it can be, and weâre trying to make it faster for them to get that sort of thing to a production ready state.
Turner Novak:
So do you know then, maybe this is kind of outside the scope of what you talk about with them, but do they just not look at those other 70% of claims? Do they just deny them because itâs just not worth it?
Adit Abraham:
Yeah.
Turner Novak:
Really? Wow.
Adit Abraham:
Well, itâs not that theyâre denying the claims that theyâre choosing to underwrite the claims that they think are relevant for them. So itâs almost like theyâre bidding against other firms, but you can think of it more as revenue leakage for them, that they are opportunities that theyâre not considering that they should be.
Turner Novak:
Interesting. Okay. So I guess one just general question I have, whatâs kind of the current state of Reducto? I know you have, some people like to say ARR, some people like to say employees. I heard you mention once that if you stacked every piece of paper that youâd processed, it was like three and a half Mount Everest. This is five or six months ago. So yeah, how many Mount Everest are you at now?
Adit Abraham:
We are probably at nine Mount Everest today. That number is changing really quickly. Even if you look at the start of the summer, like June, Julyish, we are at roughly, three times as many, sorry, five times as many pages processed per month as what we used to be at the start of the summer, and thatâs growing quickly too. We have this tracker in the office thatâs just the weekly pages that we process and it updates every 30 minutes and the team will constantly look over and itâs like a new record for how many patients were processed in that 30 minutes. Company is still fairly young. So we were incorporated September 26th, 2023, so right around the two-year mark. But we have grown incredibly quickly.
Teams about 20 people today and despite that, I think weâre lucky to say we work with what I consider to be some of the best teams in this space. That includes both the newer AI native companies. If you look at LegalTech, RV, Legora and a bunch of other folks use us there, and Rogo. Great companies that are building great vertical AI applications across horizontal platforms too, like Mercor is a customer and Scale AI and a bunch of other folks there, but also really, really large enterprises. So I mean, FAANG companies, Fortune 10 enterprises, some of the biggest hedge funds in the world and so on, across different industries.
Turner Novak:
So then how do you make money? Is it do you pay some kind of a page process type of fee?
Adit Abraham:
Itâs based on the number of pages you process and we have different API endpoints now, so... Roughly 40% of our customers actually use Reducto for two or more endpoint products. So parse their documents with Reducto and then on subset of those documents, theyâll want to do something like grab structured data. Sometimes theyâll use that structured data to then fill out a menu documents. Theyâll edit documents with Reducto as well. Itâs just the series of API requests.
Turner Novak:
And itâs kind of interesting when you just think of, you mentioned things like legal, financial services, insurance, healthcare, thatâs like 30, 40% of GDP or whatever. I think financial services, something like 10%, healthcareâs something like, I donât know, around 20% in the US. Legal is like, I donât know, surprisingly higher than anyone would realize. So itâs like itâs basically most industries rely on a lot of this stuff.
Adit Abraham:
Yeah. And I mean, those four have been a really big parts of the early traction for Reducto. You can almost think of the importance of Reducto scaling with the cost of a mistake. When you make a mistake when reading a healthcare documents, itâs not like, oops, I messed up this period versus a comma. Itâs like, no. This is somebodyâs health that youâre making decisions on and same thing in finance, insurance, et cetera. This is a universal problem, not just in those four. You can imagine all the documents that people are dealing with in supply chain.
You have bill of ladings, all of the things that you would just have people looking at manually, historically, and also a bunch of long tail cases that I just wouldnât have considered before. We have a customer that uses us to process soil analysis lab reports, and thatâs the kind of thing that we would never have thoughts to even test ourselves on. Thatâs just esoteric. But out of the box, they found that Reducto was just the best thing they could use for that and I think thatâs the value of building this as not a, we didnât want to build a financial statement parser or a health record parser. We wanted it to be, how can you read anything that comes in with the accuracy of a human reading it?
Turner Novak:
Yeah, it makes me think of maybe teachers too. I was in the orientation for one of my daughterâs classes and they write everything down and then the teacher has to look at it and grade it, and you just think of a high school teacher grading hundreds of math papers every week. Thatâs a significant amount of their time that they could be using, doing more productive things, teaching the kids better.
Adit Abraham:
We actually do have people that are building AI for education products where students will upload photos of their homework and thatâs hard because students have messy handwriting. If youâre dealing with math homework, you have equations in there. Thatâs the kind of thing that you just wouldnât have been able to do with traditional OCR that now, with the combination of both the LLMs and also frontier CV techniques, you can get there. You can read what the person had and downstream of that. You can imagine all the great things people are going to build with AI tutors and so on.
Turner Novak:
Whatâs been sort of the, I donât know, either craziest or coolest thing, use case or product that youâve seen someone build on top of Reducto?
Adit Abraham:
It depends on how you define cool because I think there are use cases that are just shareholder value. You see things and people talk about how there are repositories of documents thatâs date back a decade and now theyâve been able to digitize that and thatâs crazy. Thereâs a lot of value in that. One of the largest hedge funds in the world came to us with, they wanted to parse petabytes of data because it was all the things that their analysts have looked at historically, but the really interesting use cases to me, are actually just devoid of the practical ones. We see really interesting things come up where when the JFK files were released, those were really hard to read. I, as a human, would have struggled to read those. And we were finding cases where Reducto was transcribing the documents and it was only once I saw the Reducto parsed outputs that I was like, âOh, thatâs what that word was trying to say.â And it made sense when I looked at it side by side.
My co-founder has this really crazy story where his great-grandmother had written their familyâs history in Bengali and he doesnât read Bengali himself. And he tried uploading that into Chat products historically and it was just gibberish. It just didnât make sense. But when we first released Agentico CR, which is I think one of the biggest steps forward for us as a company, he ended up finding that that was the first time that he could digitize that family history transcription. And thatâs the kind of thing where obviously, thatâs not the focus of the company, but itâs a good reminder of how deep the problem goes, and itâs not just enterprise data ingestion, that is what we sell, but it is more generally, how does human data get reasoned on with synthetic intelligence?
Turner Novak:
Yeah, itâs pretty cool. My wife was just looking at her grandma had some handwritten recipes from over the past 60 years and it reminds me of whenever I get cards from my grandparents, itâd be written in the most intense, impossible to read cursive, and I would give it to my mom and just like, âCan you read the card? I canât read this.â Itâs just even if weâre a human, thatâs hard to read sometimes.
Adit Abraham:
Yeah. I donât know if you saw this, but there was this big historical challenge of... People were looking for humans to go through and rewrites all of the historical archives in cursive because a lot of younger people cannot read cursive anymore, our struggle too, and thatâs the kind of stuff that we want to make sure that we can just do out of the box, no matter what the content is.
Turner Novak:
Interesting. Yeah, total side tangent, but I remember back in third grade learning cursive. Iâm like, âWhy do I have to learn cursive? Whoâs going to use cursive?â And everyoneâs like, âOh, you got to learn it. Everyoneâs going to write in cursive when you get a job.â And then now, no one writes anything anymore. Weâre all like AI slop videos, just injected straight into our brain. Words donât even exist anymore. Are there any things that the product still kind of struggles with or just AI and LLMs, in general, seem to not quite be able to do yet?
Adit Abraham:
Well, a lot of our work is around, we donât really spend too much time thinking about things that are already really easy to do, like digital text for example. If you had clean, digital text in English, you wouldâve been able to OCR that well before VLMs, and thatâs fine. Obviously, thatâs table stakes for our products. Everything that we do is find the things that are challenging and find ways to help address them.
And so, youâll see all sorts of cases where, for example, thereâs no product today or there was no product, that could do chart extraction effectively because when you have graphs, like line charts, bar graphs, all of that... Bar graphs are maybe a bit easier. Line charts are incredibly difficult. If you just imagine the things that you would see in a financial research memo, a line chart could have all sorts of inflection points thatâs really hard to read, and if you wanted to turn that into table representation of the underlying data, historically, you just wouldnât have been able to do that.
And Iâm sure youâve seen sort of the, VLMs are not as bad as the tweets what sometimes make it seem, but Iâm sure youâve seen those cases where you ask a model, what does the speedometer read? And obviously, the pointer is at 20 miles per hour and it says something like 80, itâs just off. Imagine applying that to hundreds of data points to a chart, where itâs choosing where the data points is, but with some certainty that set of those data points where itâs just off by a non-trivial amount, something that data isnât useful anymore. You donât want to do anything with it.
And so, when we think about the types of things that we are doing, we tried everything that we could in terms of just training models for using other frontier models, and there was nothing that was good in a single shot image of a chart to mark down table. So instead, we started thinking of it as how do we create an environment where we can iteratively catch our own mistakes? So how do we not just make the initial representation but give it tools to be able to re-render the chart, look at the mistakes that it made, and then edit individual data points, again and again and again, until a verifier model is happy with the end chart extraction.
Turner Novak:
And then it gives the output to the end user.
Adit Abraham:
To the end... Exactly. Yeah. And itâs slower. Thereâs a clear latency cost there, but for the people that wants to be able to use that data, this is the only way thatâs possible today. And there are plenty of other things like that where spreadsheet clustering is this weirdly deep problem. I didnât know this before we started the company, but there are people who have done their entire PhD thesis on how to cluster Excel spreadsheets. And when I say cluster, I mean, you have a spreadsheet and you could have set that up any which way. You could have one massive table, you could have multiple sets of data, itâs just free form for an arbitrary number of rows and columns.
You need to know what is the disjoint sets. You need to separate that piece of information apart and thatâs a hard problem because you canât rely on formatting. The person might not have given you an outline around what they need, so youâre looking at all sorts of things like, what is the homogeneity of the data? Does it seem like thereâs a dense cluster here and a dense cluster here, and how do we separate that out? Those sorts of things are unsolved problems. Weâve done really deep literature reviews and ultimately, end up having to come up with our own techniques to be able to do it at the bar that people need because you need to be able to do this if you want to have a banker upload a spreadsheet into your AI for finance products.
Turner Novak:
So this is essentially someone doesnât design their spreadsheet in a best practices type of way, and they just have weird, their data is organized weird in weird column in rows setting, and so the computer, when it reads it, it just doesnât understand what itâs looking at. Thatâs essentially what the problem is?
Adit Abraham:
Yeah. Thereâs a lot of meaning that I think... Ultimately, these documents were made for humans like you and I to read. And thereâs a lot of meaning that is really easy for us to interpret, that when you try to codify is a lot harder. Every time you have a gap between two paragraphs, thatâs me telling you, âHey, this is a new semantic unit of information.â Thatâs just encoded and we donât think about it. Itâs just there. But if you have something like a spreadsheet, maybe they have two tables and they added one column of space, but it could also be the case that maybe that column was just supposed to be empty.
It was actually one table, but they didnât have a row of values there. So you canât hard code a value of, if thereâs one column of space, separate this out as data. Thereâs all of those sorts of things that require some amounts of semantic interpretation of the contents. It requires some amounts of just looking at that individual sheet and getting a sense of how itâs formatted. But thereâs no, or at least I donât know of a guidebook that tells bankers how to format a spreadsheet or anybody else thatâs working with Excel, and you need to be able to do that on the fly.
Turner Novak:
Thereâs actually insane, thereâs training consulting companies that will train the first year analysts. Literally, you spend the first three months on the job where itâs learning to code, but learning how to format a spreadsheet. Inputs are blue. Formulas that are linked to the same page are black. Linked to a different tab, the font needs to be green and thereâs so many different rules of... Sometimes youâll do that little, the tilde, squiggly thing, I forget what itâs called, but if you want to make it so you can control arrow jump even across empty columns. Because you know how when youâre in Excel and you jump to the end, and itâll take you to the furthest non-blank cell.
If you want to be able to navigate an entire book, theyâll add the little tildes in blank columns that you can traverse across them faster with your keyboard without using your mouse. Anyways. Yeah. There was one semester in college I helped this guy. He came to our school and trained people on Excel. It was like a crash course over a week where you paid him a thousand bucks and they trained you how to format these models. Anyways, but yeah, the banks are even more intense. Itâs literally your first month or two on the job. All youâll do is youâll go in the basement and learn how to build models. So anyways, but yeah, itâs a very manual thing. Itâs tons of rules and people still mess it up.
Adit Abraham:
And I mean, ultimately, Excel is not just used by the banks, right? It might be the case that bankers are more rigorous, but weâll have cases where the spreadsheet is actually just... We had an enterprise customer run their historical social media engagement data. They just had an export CSV that they were like, âOh, I need a language model to reason on this.â And that was just a dump. It was messy and it wasnât clearly codified with the blue formatting for inputs, all that. And youâll have things like logistics vendors where theyâll just list off the items that were purchased and still kind of do it in a free form way because theyâre just making it on the fly.
Turner Novak:
Yeah, they think of it like a Word doc. Theyâre just kind of throwing them into the cells and itâs just whatever it is.
Adit Abraham:
Exactly. Yeah.
Turner Novak:
You just hinted earlier that no one had done this until recently. Did you guys recently, are you announcing something?
Adit Abraham:
Yeah. So weâve been working on chart extraction for a while. We are about to do a really thorough write up on exactly how we achieved the results that we have.
Turner Novak:
Will that be up by the time this episode came out or?
Adit Abraham:
It should be. Yeah. Itâs actually available. Customers are using it today.
Turner Novak:
Oh, cool. Yeah, it sounds like it was a big undertaking based on what we just talked about.
Adit Abraham:
A lot of work went into it. Honestly, itâs one of those things where they went so deep into the problem that even once we hit the good enough points and we were like, âOkay, we need to do other things.â One of the engineers working on it was like, âCan I at least work on over the weekends?â I was like, âYeah, of course.â But he has been grinding through. Itâs a side project now to just have the best chart extraction possible and the work heâs done is just incredible.
Turner Novak:
Interesting. Well, thinking about that or talking about that then going beyond just good enough. I know back when you guys kind of started this, there was kind of good enough products that were out there, right? They kind of worked, right?
Adit Abraham:
Yeah, for sure.
Turner Novak:
So then, what was sort of the insight to start doing this and going back to the beginning, whyâd you guys start working on this?
Adit Abraham:
So when we applied to YC, we had built long-term memory for language models. It was, I think, the first API to do that, but it was really early. This is before GPT-4 Turbo had come out. The very first wave, chat applications was coming off the ground, ignoring ChatGPT, of course. And so, this is one of those things that naturally goes viral on Twitter because itâs cool. And to talk about memory and how itâll remember context that the user gave, but it just wasnât needed. Youâd hop on a call, we had one of those classic, you post on Twitter and your calendarâs booked out for weeks. Youâd hop on a call and youâd be like, âOh, what are your users struggling with as a result of not having memory?â And itâs like, âOh, people are barely even sending messages.â I think this would be interesting in a few months kind of thing. But one of the things that did really resonate is people started saying, âHey, if youâre managing the userâs chat history, can you also manage the files that theyâre uploading?â Like a classic just managed drag pipeline.
And we thought that that would be this low lift, we would use off the shelf tools and just add it on as a feature and it would be this fully managed platform. But surprisingly, two things. One, that ended up being the most interesting part of the platform. People got really excited by that. But two, it was the hardest part of the platform because the things that we were using, we tried pretty much everything on the market, which is keep falling apart in random ways. Reading order would mess up when you have two columns of data. It would read left, right instead of column by column. Youâd find all sorts of cases where the table wasnât parsed correctly and so on.
And so, it just became this thing that we were spending so much time building custom logic around. We eventually had to start training our own models around to segment the layout, all of that. And where Reducto is today wasnât really us deciding like, âOh, this is a $10 billion opportunity. Letâs pivot into it and do everything with it.â Actually, it started as a technical marketing blog where we had this really ugly Streamlit app. You just drop a document and we would draw boxes on the document. I cannot exaggerate enough how simple it was, but that just immediately got a ton of interest. We started seeing a lot of founders for, but at the time, were early stage companies but are now kind of household names, talking about how, hey, this is better than what Iâm seeing from companies like Textracts and others, even folks that were focused on the space. And we would hop on calls with them and realize that they were going through the same process that we had.
They were spending a ton of their time post-processing outputs, building custom fallbacks. It was just this thing that was holding them back from doing the things that they really want because if youâre a legal tech company that parses contracts, itâs not the parsing that youâre interested in. Itâs the intelligence on top of contracts that your customers actually care about. But if you parse it incorrectly, thatâs the bottleneck that holds you back from being able to give good insights to the lawyer. So we thought of it as, what would it mean for us to almost be the ingestion team for our customers? What if we could actually solve those last mile problems such that ideally, they shouldnât even be thinking about PDF processing? And thatâs how we went down that road and very quickly started seeing really, really intense adoption. I think we were talking on email and I mentioned we ended up getting a Fortune 10 customer as a full annual enterprise contract as a team of two.
Turner Novak:
Yeah, so tell me about that. How did you land that? I think you guys were still pretty early on. The product was pretty new at the time.
Adit Abraham:
Yeah, I can send you a screenshot of how ugly our website was. It wasnât this nice enterprise ready professional page. It wasnât meant to convert. It really was a super simple text tagline and the important thing was we had, again, a really ugly front end application that would at least show you what the product does. So from day one, we had a playground where you could test Reducto for yourself because we knew that it was a really noisy space. Itâs been around for a while. People have seen decades of PDF processing companies and even today, every week you hear about something, not just in our space, across the AI landscape, you hear the new best or the first AI agents for blank, and itâs actually the 20th.
And so, we wanted people to just be able to kind of see the proof is in the pudding when we say that weâre the most accurate. And I donât know exactly where they found this from. I think they mightâve seen a LinkedIn post or something like that, but they decided to upload one of the, what I call gotcha documents. Everybody has some set of documents that they expect to fail. Itâs like a thing that theyâve been trying to solve and it just hasnât worked. And that, out of the box, ended up working in the playground.
Turner Novak:
Thatâll get their attention pretty quick.
Adit Abraham:
Yeah. It was like, okay, interesting. Letâs have a conversation at least. So they came inbound. We had a conversation with them and obviously, it was still a multi-month process because this is a very technical company and they had a whole team of people working on intelligent document processing, so trusting a two-person startup just almost doesnât make sense. Borderline feels irrational, Iâm sure.
Turner Novak:
Yeah. Theyâre like, âWho are these kids? What are they, lying to us? Is this fake?â
Adit Abraham:
Yeah. Honestly, there was a lot of skepticism when we first got started, and I understand itâs because the company had just been founded. We hadnât even announced any sort of a Seed Round at that point. And so yeah, there was a lot of just intense back and forth, just stress testing everything that weâd done. At some points, they had 15 of their engineers meet with us for seven or eight hours across the day.
We were just in a conference room, stress testing everything about Reducto, like what it could do, what it couldnât do. And I really donât think it was like... There are sales lessons that I have coming out this, but it wasnât, oh, weâre incredible at sales and thatâs why we landed this customer. I genuinely think it was just we had an incredible products from day one. We spent a while working on it before we even released it as a product and the product kind of spoke for itself.
Turner Novak:
You mentioned you just kind of made a post, I think it was on the YC internal forum, so itâs private. So, not even a big viral thing. Thatâs kind of the antithesis today is everyone needs to launch with a launch video. Right? Itâs kind of the meme. You guys did a private, non-viral opportunity just a post with some pictures and screenshots and stuff, and it was basically just you showed that you solved this problem that people had and people resonated with it, and it was really that simple?
Adit Abraham:
Yeah, at least for the early points and obviously, it didnât stop at that internal YC post. We started posting a lot more publicly over time, but for a while, we didnât even... For the launch video category of companies, I think there are cases where you just want to go big really fast. You want everybody on your platform. Maybe youâre a B2C founder and you need that. For us, we actually gated demand a bunch. We didnât have self-serve onboarding, which for an API product, is kind of weird, but we kept that true, I think, until a few months ago. So for probably a year, you could not sign up for Reducto without us onboarding you. And part of that was we, from day one, have been adamant. The only reason the company has a right to exist is if we are the best product in this space. That needs to be true.
We didnât want to be the 200th PDF processing company that fails to deliver on the promise. And so we can talk a lot about this. We just would say no to people if we didnât think that we were at a point where we could solve their individual use case. Itâs such a deep problem that you canât solve everything from day one. And so, if you came to us with a chart extraction problem two years ago, even though we can do it now, historically, we wouldâve said, âHey, we donât think weâre the right vendor for you. Weâll do it someday, but weâre not doing it yet.â And all sorts of things like that. To start, all we said we would do is the very first thing that we decided to train our own model for, was we wanted best in class layout detection and understanding. And so, we spent a lot of time building that and just being incredible at layout understanding.
And then from there, we decided we wanted to be incredible at table parsing and table detection. And so we spent a ton of time on that and these are broad horizontal problems. If you have great table detection, it ends up applying to a bunch of different industries and same with layout and the other things we did. But it was this iterative process of growing the scope, but for every individual stage, we made sure that we felt like we were incredible at what we were doing. And I know you mentioned this earlier of there were pretty good solutions available when we started, I think one of the closest comparable companies to Reducto when we first got started, along this idea of unstructured data for LLMs, had this thesis around you can upload any file to this platform. They had 35 different file types.
Turner Novak:
I think itâs 64 now, isnât it or something like that?
Adit Abraham:
Something like that. The number grows week after week.
Turner Novak:
Yep.
Adit Abraham:
And a lot of our early customers were actually churned customers of theirs. And I talked to them, I was like, âWhat didnât work?â Whether they needed us to support that many file types because there was no way we were going to do that on day one. And they were like, âNo, actually, we only really have two to three types of files that really matter. This is the 80/20 or 95/5 in our case, and we just need those done incredibly well and we canât spend all this time with okayish pipeline for things that we donât even care about when the things that we do care about, donât work.â And that was really important. Itâs like, it was okay that we werenât competing on number of file types. We were just competing on for the people that needed PDFs and images, great, Reducto, itâs a solution for you. And then gradually expand it from there.
Turner Novak:
So itâs almost like a 80/20 rule, but it sounds like it was more of 95/5 rule or whatever, where itâs like people maybe only wanted a couple things. It was the vast majority of actual customer demand versus fancy marketing or like, oh, we can do everything, but the whole going super deep versus being super wide.
Adit Abraham:
Yeah. And this is actually something thatâs taking a step back from Reducto, I just generally feel is true in startups. I think you see a lot of the foundersâ DNA in the ethos of the company, for any company. And for some of these competing companies, if you look at the foundersâ background, itâs like a go-to-market and marketing and sales background and that has its merits. Iâm sure it helped a bunch for the early days, whereas Raunak and I really started off more on the product side. Raunakâs been doing research since quite literally, he was 12-years-old. He published in high school when he was 14 or 15 as a first author on a computer vision paper. And so, when we came into this, it wasnât the case that we were incredible at sales. It wasnât a case that we knew a ton about marketing, but what we did know is, we can build something incredible that once people use it, we can still understand it. And everything else has been a learning journey along the way.
Turner Novak:
Yeah, I think you mentioned just talking about Raunak for a second. I think you were in some kind of class at MIT and he was a freshman teaching the class or something like that, if I remembering that story right?
Adit Abraham:
He was learning assistant for, it was my first graduate course and I remember I had a ton of imposter syndrome because I was a junior in my undergrad and most of the people in the room were PhDs. Thereâs a course on meta-learning, so teaching models to learn. And yeah, Raunak the... He did the walkthrough of how we would approach the first --sets. His approach to the problem was the guide for solving it. Thereâs a whole different rabbit hole that we can talk about between the professor of that class, Raunak ended up getting that professor fired, the really viral Twitter post. But yeah, thatâs a deep rabbit hole.
Turner Novak:
Oh, no. Hopefully... Oh, man. I donât know if weâre going to go down that route. Well, right now, if we had an extra maybe 15 minutes, but... Oh, man. Thatâs crazy.
Adit Abraham:
The short summary is that that professor had published a paper called, GPT-4 Can Solve MIT. This was one of the really famous papers around when there was all this language model hype. And then Raunak did breakdown of the paper talking through all the reasons why the results were invalid and all of this, and this was when he was a freshman or sophomore. And there were just crazy aspects to it where there would be a multiple choice question and the way they were evaluating was they would have GPT-4 being prompted in a four loop, again and again and again, until they answered the correct answer in the multiple choice. And then it would like, oh, it got a hundred percent on the score. So it just was nonsensical and in the process of Raunak doing that write-up of all the reasons why it wasnât good research, it ended up coming out that the professor also hadnât gotten permission to scrape a lot of this data. And so it just became this bigger thing that blew up and eventually, the professor got fired from MIT.
Turner Novak:
Wow. Thatâs a pretty big deal. Obviously, you got to get fair usage rights of the stuff that youâre using.
One slightly different topic, but I donât want to miss it. You mentioned that you guys, kind of this whole concept sort of pivoting the product. I think you kind of made it, it sounds like a pretty bold pivot at the time. How did you decide to do it? Because I know youâve said, Iâve heard you mention in the past, you were pretty cognizant of, you donât want to kind of be aimlessly wandering around trying all these different things. How did you approach the process of like, this is our specific sniper shot, organized pivot that weâre going to do?
Adit Abraham:
Yeah, I think one of the worst things about pivot hell as a founder is that there are so many surface level interesting opportunities that have a one layer deeper for whatever reason, are hard or not viable, whatever it might be. If the way that youâre approaching pivoting, and weâve done this at some point before we landed on Reducto. If the way that youâre pivoting is, you pull up a Google Doc and youâre brainstorming all these cool startup ideas and youâre going back and forth of like, âOh, shoot, thatâd be so sick.â You end up in this horrible position where you land on an idea that you think is incredibly insightful. Itâs like your eureka moments. And then as soon as you start Google searching, you find that somebodyâs tried it before because markets are smart. If you thought of it in a brainstorming session, somebody in that space has probably thought of it too. And then youâre in a horrible loop where you never make progress in any given thing. Youâre just bouncing from thing to thing.
So we knew we didnât once to do that because all of the interesting work is, you have some sort of insights, you find the reason why the insight is not as obvious as you thought or theyâre blockers and not just being free money or a market inefficiency, but then you go one layer deeper of, okay, other companies exist. What are they not doing well? What is the one step deeper reason why something could be better? And so I think the space that weâre working on, hit a few different check boxes for us that it made sense to do. One is, we felt we had an unfair advantage in computer vision. Like I mentioned, Raunakâs been doing computer vision in particular for basically his post adolescent life. And so, it just seemed like a space where we could offer a lot of expertise. Two is, we had spent a lot of time around this area, so we understood the problem well. The way that we actually started working together is we competed in the Anthropic Hackathon, where they first released Quad 2.0.
We got early access to Quad 2, before it was publicly released, and we won that hackathon together. And one of the things thatâs... One of components of what we built there was the ability to upload PDFs and we saw the details of where foundation models struggled here. And three, I think the two steps removed version of the company was really, really exciting. On face value, working on PDF processing felt like not sexy problem, but looking at it as this broader, what does it mean for language models to interact with human data? What is the scope of that? Clearly, even back then felt like something that had to be huge. If it wasnât going to be us, somebody would figure out this problem and that was the trifecta that was necessary of at least trying to go deeper in it. And we committed at the start of YC that we wouldnât pivot during YC, at least not a meaningful pivot. We wanted to see YC through with really pushing on one thing with a lot of focus and dedication for that thing and that paid off.
Turner Novak:
Yeah, it sounds like itâs almost figure out a core problem or trend or customer base to serve. It sounds like what you described doesnât quite fit in any of those buckets, but it sort of does. But itâs kind of like find the one thing and then you can kind of bounce around in it and then until you dial into the specific part, and then it sounds like you kind of expanded from there, once you actually figured out the core initial thing that people wanted.
Adit Abraham:
Yeah, and I think our space was especially interesting because people... Ignoring the AI market, imagine it didnât exist, thereâs already billions of dollars spent per year on document processing. Thereâs an existing markets where if we were trying to build a marginally better solution, thereâs probably some money to be made. Itâs not an exciting company, if that was the market. It wasnât us trying to create something that new that was exclusive to language models. To us, it was this two-sided thing. There are legacy use cases that we can just be an order of magnitude better at, but also as this new generation of companies are being built, we can be core infrastructure for them to build better products.
Turner Novak:
And I think you, Iâve heard you say before, I think this is public, you got to about a million in revenue within a few months. So it sounds like you had some startups who were using you, obviously. You signed a couple big customers. I think maybe we could talk now about the fundraising journey you guys have sort been on. I think you mentioned two years old, raised a total of, I think you said 108 million? Maybe Iâm not remembering right. And youâve only burned a million throughout the whole course of the company, something like that, yeah. Maybe between now and when we publish, youâll have hired a couple more people or something, and momentarily maybe youâll sign a new deal though. Maybe itâll go back up. So just in terms of fundraising, whatâs that been like over the past two years?
Adit Abraham:
Yeah, weâve been really lucky in that I think at each rounds, weâve gotten to partner with whoever we were most excited with or about. And because we werenât raising as a, shoot, weâre about to run out of runway, sort of position, it also meant that we werenât approaching funding rounds as this needs to happen and more so as a opportunistic, there might be room for us to really accelerate the company as alongside this person. So we announced our Seed Round, I believe in August last year, raised a Series A. Few months later and most recently, a few days before this podcast was released, weâll be announcing our Series B from Andreessen Horowitz. And at each one of those stages, it really has been a question of, what is the next phase of the company?
I remember when we raised our Series A, we were four people, I donât know how many Benchmark Series Aâs are done at that team head count, but they saw across their... Some of the Benchmark portfolio companies were already customers of Reducto before Benchmark invested and they saw how much those companies loved the product. And that, I think, speaks really, really strongly. The number one thing that has always mattered to us is not VC interest. Itâs how strongly do our end customers feel about what weâre doing? Does it feel like something that is just apathetic part of their stock that they could just replace at any points or does it feel like something that they are truly excited to build alongside? And I generally believe that funding is almost like after effects of the first thing being true, whereas I think a lot of founders try to look at it the other way.
Turner Novak:
Yeah, so it sounds like, I think, you probably only raised a couple million in the Seed. Youâve only burned a million. So in theory, you didnât actually need to raise any money. Why did you think you needed to raise a Series A?
Adit Abraham:
So there are a few things. Youâre right that when we raised a Series A, we hadnât even spent our YC Funding at that point. We really liked Chetan at Benchmark. Donât take this the wrong way, but itâs not often the case that you meet with a VC and the way that theyâre thinking about your business is genuinely insightful. And I told Chetan for months that, âHey, weâre not going to raise. This doesnât make sense.â Weâd received a lot of preempted interest outside of Benchmark before that as well and we just kept saying, âNo, this doesnât make sense. Weâre not going to raise now,â and so on, but we would just continue to have dinners and every dinner, it just ended up being a really productive conversation where over time, we started to really think about it not just as cash in the bank thing, but also as, we want this person on our board. We think that this would help in the next phase of the company.
And obviously, I think Iâve heard the warning stories of raising money when you donât need it. To me, a lot of that comes from being undisciplined with the money that you raise. Thereâs a great way to spend VC dollars. Theyâre a tool like any other. And weâve had the luxury of anytime we want to work on a frontier model, we donât think in terms of compute costs. Weâve never told our ML researchers, âHey, you have this cap on the number of GP hours.â So weâre purely thinking in terms of the end outcomes that we want and thatâs the luxury that we get as a result of VC dollars. But the thing to me is, we didnât want to see raising money as a mandate to immediately spend the money. Weâre obviously ramping up our spend, especially as weâre scaling go to markets, but we were on the same page as the investors that preempted us about what the next year of the company looked like and at that point, it just made sense to raise because terms were also sensible.
Turner Novak:
Heâs another Benchmark Portfolio Company, Bobby DiSimone at Pomerium. He was telling me one of the board meetings they had, they were accidentally profitable and it was kind of a joke that theyâre like, âYou got to be spending money. Why are you not burning cash? We gave you a bunch. We donât want you to be profitable right now.â But thatâs good. I mean, itâs good that you have a real business that makes money versus... And solve problems for customers, being a vehicle that transfers LP dollars to the landlords in San Francisco.
Adit Abraham:
Even now coming out of the Series B, I remember I sent out... Our first investor update following the Series B, I sent a note saying, âHey, weâre really ramping up and I do expect burn to ramp up.â And so we started being more aggressive with spending, but also, the following month, even before us announcing the round, ended up being the most six figure contracts that weâve ever closed. And so, our spend went up, but our net burn didnât. So it kind of looks like we are just this flat net burn curve even though thatâs not intentional, that accidentally profitable line resonates.
Turner Novak:
Thatâs a good skill though because when youâre a public company, theyâre like, they want you to give these exact targets and being able to continue to make it look like, even though every company is kind of a shit show, everythingâs always crazy. They want it to look from the outside like, âOh, these guys are so good.â The CFO is so good at guiding the street and hitting the exact target and look like itâs super smooth. So you guys got it down two years in. Thatâs a good sign.
Adit Abraham:
With the no finance background, I guess.
Turner Novak:
Yeah. Well, you got all those finance customers, maybe youâre learning something, talking to everybody. So you mentioned that it took you about 48 hours to do it. What was the process of doing a Series B? I think, donât people usually people big long process, tons of meetings, et cetera. What was it like for you guys?
Adit Abraham:
So we very much did not want to do that. One of the problems with having 20 people working at the scale that the company is at today is that I really think that every hour of time really matters. And so, we didnât want to distract ourselves from the things that really matter of sales, products and customer support, to just go on like a goose chase. So, well before we decided to raise a round, I opportunistically would take meetings there as an investor or actually, as of this podcast, this will be public. Iâd known Jennifer since our Seed Round and really enjoyed our conversations. Weâd kept in touch across a year plus, and so we had a pretty strong relationship, which I think maybe is not always true when people start a process. But when we started or when we came to a decision on whether or not it might make sense to raise a Series B, this is something I went back and forth on with a lot for our existing investors.
We just came into it with a clear mindsets of, weâre going to run a very tight process. This is not going to be put the company aside for the sake of chasing around. And so, we went with the two firms that we were most excited about. We have a short list of five firms that we would want to raise from. On Sunday, I kind of just told the two that we end up actually pitching to, âHey, somethingâs happening. If youâre still interested, nowâs the time.â And then the following 48 hours were really, really intense. We obviously werenât officially planning to do this. So we hadnât done really thorough and made a data room, but needed to put one together because they needed to see everything about the company. So it was like the scramble of exporting all of our Stripe data and all of that kind of stuff. That was my full Sunday. I probably called the two investors few dozen times in that 48 hour window, where it was like every 45 minutes, weâd be on the phone because everybody was trying to condense everything.
The following day, we did the two partner meetings, we got two term sheets and both of these firms were firms that we were really excited about. So it wasnât like if we get a term sheet from one, weâre just immediately saying yes. But at the end of the day, I think there are a few things that matter in the funding ground and I donât think that valuation is the number one thing there. So we werenât too concerned about just artificially bidding up just the raw numbers in the rounds. The things that we really did care about is making sure thatâs a given firm was right for exactly what we were trying to accomplish in the next 12 months. And especially in the spirit of what I mentioned of, Reductoâs had incredible traction without a sales team, primarily, almost entirely from word of mouth with our customers talking about the products. Nowâs the time when weâre ready to really work with the broader range of customers that could benefit from Reducto, and Andreessen felt like the right firm for that.
Turner Novak:
And it sounds like youâre going to start to hire up a little bit more on the team. I think you mentioned thereâs about 20 people. You mentioned the first AI researcher that you hired? I know itâs an interesting story. What happened there?
Adit Abraham:
So Yifei is incredible. We knew about Yifei from very early on in our company history. So just for context, he was this PhD student at Purdue. He was doing his entire PhD on vision models for document processing. Probably one of the best experts in the world at exactly what we do. And itâs rare that you find that sort of fit.
Turner Novak:
Yeah, itâs probably the most boring PhD to think of like, âI want to do a PhD on this,â but the perfect fit for you, for Reducto.
Adit Abraham:
Itâs rare though, you find somebody like that, but he knew all of the nuances of why the problem that we were working on matters because he was spending all day thinking about it. And so, as a side project from his PhD, this isnât even the frontier research that he was doing, he had open source models for this space and they became number one on Hugging Face. I remember we even had a competitor who has since shut down whose whole product was built on Yifeiâs side project open source work. They were wrapping that model. Great work, honestly, especially considering it wasnât his core focus. And so, we spent a while just trying to get on the phone with him. Raunak had DMed him, and I think generally speaking with somebody like him, thereâs always a ton of companies that wants to convince them to come on board. I DMed him again two months later and eventually just got him to take one 30-minute call.
Turner Novak:
Howâd you convince him?
Adit Abraham:
It was just us texting on Twitter. So it was a Twitter turned into Google Meet. And I think even on that first call, you could kind of see that a lot of the ways that we were thinking about the problem, what it could be, were very just much aligned, but he still had some reservations because we hadnât even announced the Seed Round. We were two person companies still, just made our first founding engineer hire. And so we decided to tell him that, âHey, actually, we have some meetings in Indiana and I know youâre based in Indiana. Can we just grab lunch on the way?â I only found out recently that he actually believed that we were in Indiana for no reason.
Turner Novak:
Amazing. The most random place for you to be.
Adit Abraham:
Yeah. Heâs a very sincere person. So yeah, we told him, âHey, weâre going to be there. Can we get lunch?â And as soon as he said yes, we took a red eye over, flew to Indiana just for this classic brushing in the airport bathroom, and then just going directly. And it wasnât just lunch, we ended up spending seven or eight hours together. We were just walking around the Purdue campus talking about everything that we were doing. Partially, it wasnât even necessarily us interviewing him, it was also him interviewing us, really digging into why we made the decisions that weâve made and all of that.
And I think the thing that really sealed it for him, going back to what I mentioned, our customers choose us because the product is best in class. Yifei had offers from some of our competitors as well, and he tried everyoneâs product and ended up seeing the best results from Reducto. And again, there were competitors that were larger teams and the same way it applies to our customers, I think he just saw this as a place where heâd be excited to build that sort of state of the art. And so yeah, weâre lucky to have him on the team. Heâs been on the team for quite a while now or quite a while in the context of a two-year-old company.
Turner Novak:
Yeah, percentage of the time the companyâs been alive, yeah.
What else are you guys hiring for? We can plug it obviously right here really quick, but for people listening to this right now, thinks it sounds interesting, who are you looking to join the team?
Adit Abraham:
Yeah. We are hiring across the board where on engine research, that is a role where we are always looking for exceptional people for all the reasons that weâve talked about today. These are really hard problems and we need exceptional people to be able to solve them. We also are hiring for design as we build out the platform for Reducto, but especially the thing that Iâm really excited about is, we are newly hiring for Go-To-Market.
For a long time, we would get salespeople reaching inbound and we just werenât ready to transition away from founder-led sales. But even in our early hiring, our first AE in his first four weeks, Iâm including ramp-up time here, closed a six-figure deal from zero initial discovery call to contract signed, and he was talking about how in his entire career so far, he has never had a deal cycle move that quickly for a greater than a 100K deal. And that, to me, is the signal for okay, we clearly have way more demands than we have capacity to sell. So weâre growing our AE team, hiring a head of marketing and so on.
Turner Novak:
I think you told me before you got to about 5 million in ARR before you hired him, before you brought anyone on the sales side, and you described yourself as not a sales person. Iâm listening to you talk, doesnât sound like youâre selling me anything, like the classic snake oil salesman. How did you get good at sales? Whatâs a general, we can unpack this a little bit more, but at a high level, how did you learn how to do this?
Adit Abraham:
I know that one of the things that sales leaders always say is, âYouâre not supposed to demo on the first call. Donât do it. You should just spend your time doing discovery.â Thatâs been historical advice that Iâve been told many times because I guess the nature of the advice for a long time was you need to spend a lot of time understanding the customerâs contacts before you come to them with a nice polished demo. And part of this also is the traditional structure. If you have a non-technical AE that then needs to bring in an SC for the demo, that sort of work. I think that the benefits that founders have when theyâre selling is, founders know every little nuance of the product on a really deep level. And to me, it would be a shame if that wasnât captured, even in the first call because the way that Iâve kind of mapped out sales, I know thereâs traditional sales milestones, but to me, thereâs really only two phases that matter.
The first is the inspiration phase, getting the prospect to a point where they want to buy, and as soon as that is done, the job of a salesperson becomes facilitation. Itâs you going through the steps of walking through procurements, talking to the rest of the team and so on. Iâm being redundant here, but you get what I mean. And so, the thing that we have really optimized for is, I try to get people to that inspiration phase as early as possible and oftentimes, thatâs in the first call. Weâve had cases where people will bring in C-Suite on the second call for really large companies just because the first call was something that they were just buzzing about. That kind of thing is something that I think, if I came from a traditional sales background, I wouldnât be approaching it this way, as a benefit of being an amateur in the space. With that being said, thereâs plenty of things that I didnât know about. Iâd never gone through a procurement loop, and thatâs something where no amounts of product expertise even matters. Thereâs just traditional knowledge that you need.
Turner Novak:
Oh, yes. So I guess for somebody whoâs never gone through that, can you just describe why thatâs a big deal, like the shock that you get?
Adit Abraham:
So for people that havenât done sales before, thereâs two components or every large enterprise will have the sales process that youâre going through to convince the end buyer, the person thatâs using your product. And that, if youâre aligned on product and what it does, is hopefully the easier one of the two. If the value of the product is there, people should be aligned. The second part is that enterprises have so many things in place to prevent bad purchases. That includes their security review team. Thereâll be dedicated teams that are just there to make your contract as difficult as possible. Theyâll negotiate everything. Itâs not just pricing, itâs the renewal terms, itâs the time that they take to actually send you the payment for the invoice, all of that kind of stuff. And that can be, to me, that seems like some, or at least when I came into this, I wouldâve imagined that would be a one-week process. In practice, enterprise procurement can be a month plus process, if not longer in many cases.
And so, that sort of stuff I ended up just needing to lean a lot on our investors, like First Round was incredibly helpful here. Emory at First Round actually guided me through the process of our first procurement loop and it just became something that we got better at over time with more and more reps. But the challenging thing I think as a non-seller is similar to how as a founder, youâre at a disadvantage against VCs because VCs are seeing thousands of deals, and you donât even know the scope of what you can really negotiate unless you have a lot of founders you can lean on. Procurements is seeing tons of deals done and the moments they sense weakness, some procurement people will just jump in and theyâll be like, âYouâre a two-person company, we need to access deal unless you can meet this price.â Things like that where that just arenât true. Thatâs as a first line founder can sometimes fall for.
Turner Novak:
Someone described it to me as basically like selling software. Thereâs no cogs, so the price can be a dollar or a million, right? It can be anything in between. Itâs just, are you delivering enough value to make it worth them paying that? Thereâs a 10X ROI or whatever, I donât know what the number you pick, but you got to make sure itâs actually valuable and then theyâll justify paying whatever this pretty big price tag is at the end of the day.
Adit Abraham:
Yeah, and I think the thing that Iâve realized over time is when I first got into it, procurement was painted as this antagonist. And I think that on some level, the incentive structure is there, but for a lot of our enterprise customers, Iâm actually on really good texting terms with their procurement leaders now because the same applies to them, right? Iâm sure they deal with vendors that are really difficult, vendors that donât live up to the promise of what they put in a contract.
And so everything that a seller might say about how difficult procurement can be, procurement can talk about of how painful vendors have been for them. I think if you approach it with that sort of empathy of they are trying to do whatâs best for their company, you find that theyâre not actually trying to screw you over, generally. Theyâre just trying to make something work that works on both ends and people are willing to be reasonable if youâre reasonable.
Turner Novak:
One of my portfolio companies, his name is Chris Hladczuk. He has a company called Hanover Park. I donât know if youâve come across them.
Adit Abraham:
Iâve seen them.
Turner Novak:
Yes. So they do, itâs like fund admin for investment firms. I donât know if they use Reducto, but maybe they should because they do a lot of document ingestion. Maybe Iâll text them after this. But one of the things he told me that works really well is, just you need to immediately get people on a second channel, right? You need to be able to have, maybe you get introduced on email, but you got to get their phone number and just have just a different way to talk to them. So email maybe itâs a more semi-formal follow-up, and the text is, I donât know, a funny meme related to their product or something.
Just another way to stay top of mind versus a, âHey, just following up. Did you see the thing? Or do you have any questions?â You kind of need a reason to stay on top of people thatâs not bugging them that actually kind of seems natural and like, âOh, just thinking of you. Love you guys so much.â So thatâs something Iâve always kind of kept in mind. Itâs like, immediately get their number. I think thatâs actually something Chetan does too at Benchmark when Iâve learned. Itâs like, get their number, get in their calendar, plan the dinner, plan the next dinner after that, you just get in their flow almost.
Adit Abraham:
Heâs really good at that. Itâs like what we were talking about earlier of me telling Chetan, âHey, weâre not going to raise. Iâll go to dinner, but no expectations for dinner.â
Turner Novak:
One last question I had for you. So with the models kind always changing, I mean everythingâs changing. I think the classic in AI is, you canât make plans because in a week, everything changes. Iâm sure by the time we publish this, everything we talked about will be irrelevant two weeks later. So what does that look like behind the scenes on your end? Just how do you stay on top of it?
Adit Abraham:
I think one of the mandates for Reducto is not even just train great models. Itâs help our customers not have to think about PDF processing. And so generally what that translates to is, anytime a new model comes out, like in GPT-5 or someday like a Gemini 3, our customers shouldnât have to think about, âOh, do I need to swap this in for some subset of my documents?â All of that. We should be doing that for them. And so, weâve put a ton of time into just being really rigorous thatâs evaluating on all sorts of different subtests.
We will swap in models, post-training models really quickly whenever we find that, hey, this is better at, letâs say, itâs checkbox detection, whatever that subtest might be. But the other side of that of just making sure that your product is always at the forefront is the most interesting thing as a founder in this space is you end up coming across new opportunities that just werenât possible before, more often than I think you wouldâve pre-language models. I remember we wanted to do document editing around this time last year.
In 2024, we were curious of how can we fill out documents that are scanned, which is a really hard problem. Itâs almost like the reverse of everything weâve talked about. You need to figure out what the empty space actually translates to in terms of putting something in or if thereâs a square, needs to figure out, oh, this is a fillable checkbox, all that kind of stuff. And that just wasnât possible to do well with models that existed then. And we tried a bunch of different things, but it just wasnât a product that we wanted to ship because it wasnât good. It didnât live up to the promise.
And it wasnât until recent advancements that that did become possible. We released what I think is the first truly horizontal document editing API and it requires everything. Weâre using reasoning models, weâre using Frontier VLMs, weâre using traditional CB. And all of those put together is the first time that you can do something like this. And I think that will continue to happen, maybe on the order of when this podcast is released, maybe on the order of a few months. But I think you just constantly have to be looking at not just what is the problem today, but what will people need in the next few weeks? What is now possible that just wasnât possible a few weeks ago?
Turner Novak:
Interesting. I mean, and it begs the question maybe last point is just, what is the future of just all this look like, like when you just fast forward to the end? I donât know if youâve thought that far, but whatâs it all look like?
Adit Abraham:
The end state of PDFs and documents is, I think, ultimately, these documents are proxies for human work. Theyâre generated by humans as an artifact of the work that they did. Thereâre the source through which people get information for their work. And the first wave of AI was sort of like synthesis and summarization tools. It was, you had some corpus of information and you want it to be able to ask questions from it. You want it to be able to pull insights, whatever that is. But the place where people are clearly headed, as you see more of these AI employees and agentic workflow products and all of that, is people are doing that end-to-end.
Itâs not just about answering a question that a person answered or asked. Itâs, how do you fill out the end-to-end process? We have customers that do security questionnaire automation and theyâll just take the context about the customer. Theyâll parse all that with Reducto. Theyâll extract the questions from the questionnaire and theyâll actually fill out the questionnaire for you so that you just upload that completed questionnaire. That sort of thing, I think, is really interesting because thatâs the points where itâs not purely a tool to help you do small subtests. Itâs a tool that takes things off your plate.
Turner Novak:
Yeah, itâs kind of like when you think about this whole agent to agent promise that we have. Itâs just instant transfer of tons of information. Well, this has been a lot of fun. Thanks for coming on the podcast.
Adit Abraham:
Yeah, this was great. Thanks for having me.
Stream the full episode on YouTube, Spotify, or Apple.
Find transcripts of all other episodes here.