# Why Datasite Stopped Building Its Own Parsing - Adit Abraham, Reducto & Raj Bakhru, Datasite

Adit Abraham, Reducto · Raj Bakhru, Datasite

[Talk page](https://ai.engineer/talks/0PqGT2L6cZM-why-datasite-stopped-building-its-own-parsing) · [Video](https://www.youtube.com/watch?v=0PqGT2L6cZM)

## Why Datasite Stopped Building Its Own Document Parser

When BlueFlame traced a wrong answer for a private equity customer back to a misread pie chart, its engineers saw how much could go wrong before a model answered a question. Datasite later credited Reducto’s parsing with a 22% improvement in recall on its internal deal-room benchmark.

### From building trading systems to working on the deals themselves

[0:12](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=12s)

Adit Abraham, co-founder and CEO of Reducto, opens with a figure he wants Raj Bakhru to confirm. Roughly 60% of transactions worth more than $100 million go through Datasite, where Bakhru is global CTO. Bakhru confirms it. Abraham’s interest is how the engineering decisions look from two positions Bakhru has held. He helped found BlueFlame, an early vertical AI company serving mergers and acquisitions, then joined the company that already hosted the documents for much of that market.

Bakhru came to M&A by way of software development. At Goldman, Highbridge and Kepos Capital, he built research and trading systems for quant funds. He then started a cybersecurity firm and sold it to a private equity-backed company, where he ended up running M&A and corporate strategy. By the time he started BlueFlame, he had experience both building systems for finance professionals and doing transaction work himself. The founders saw an opportunity in the volume of unstructured information people had to interpret to complete a deal.

BlueFlame spent roughly a year to a year and a half building before Datasite approached it about an acquisition. Bakhru and his co-founders were financing much of the work themselves. Building their own parsing, document processing and ingestion pipeline was one way to conserve cash. It worked well enough to put a product in customers’ hands. Those customers would give the engineers much more demanding examples than the ones that had convinced them the pipeline worked. BlueFlame had already started working with Reducto while it was still independent.

### Putting question answering inside 55,000 deal rooms a year

[2:12](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=132s)

BlueFlame had about 50 people when Datasite acquired it. Its first assignment was to bring AI into Datasite’s virtual data rooms, or VDRs. When a company is for sale, bankers use a room to share the transaction documents with prospective buyers, including private equity firms. Diligence providers also use those materials. A user submitting a question inside a room wants an answer grounded in the documents for that particular sale.

Bakhru puts the scale at billions of pages of content, more than 500,000 monthly active users and about 55,000 data rooms a year. Users were asking for AI features regularly, and the board wanted them too. To answer questions across those rooms, the team needed to parse the files, ingest the extracted content into a vector database and build its graph-based retrieval-augmented generation system, or GraphRAG, on top. That sequence made extraction quality an issue for everyone working downstream. A query could reach the right document and still yield the wrong answer if the parser had misread the relevant figure.

There was also a constraint on where the processing could happen. Datasite had historically kept the contents of its data rooms inside its Azure environment. The files held material nonpublic information about transactions. Clients had regulatory obligations that carried through to Datasite, and Datasite had its own certifications, including ISO certifications, to maintain. Bakhru’s team spent time with Abraham’s working out how to use Reducto while keeping the content protected. Choosing a parser on quality, speed, recall and cost was one decision. Establishing where it could operate on live deal documents was another.

### Tracing a private equity firm’s wrong answer back to a pie chart

[4:41](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=281s)

BlueFlame’s data science team benchmarked the extraction and ingestion tooling the founders had built. Its most instructive debugging material came from private equity customers who sent back questions the RAG system had answered incorrectly. The support case arrived as a bad answer. To understand it, someone had to work backward to the source and determine what information the answering system had actually received.

In one case the source was a pie chart. Reading it required associating a shade of blue in the legend with the corresponding slice. Even using what the team considered the best available vision model, the model read that association incorrectly. Bakhru says the same model would have read the same data correctly as a bar chart. The financial question and underlying numbers could stay the same while the representation changed which fact the system extracted. Improving the prose of the final answer would do little for a value already attached to the wrong category.

Finance workbooks produced a different class of support case. Merged cells made it difficult for a model to interpret how a sheet fit together. Bakhru points out that the workbooks can be hard even for a person to understand. And making sense of the layout in one workbook addressed only part of the job. The team also encountered binary Excel workbooks, macro-enabled files and template files. Each was another form in which a client could reasonably expect to submit financial information.

### Drawing the supplier boundary at the output of extraction

[6:11](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=371s)

For Bakhru, the cost of owning the parser became apparent in the support queue. Someone had to investigate the chart, work out which model read that type of visual reliably and then take the next case involving a different chart or spreadsheet. The original decision to build had saved money when cash was scarce. Maintaining that decision committed a limited group of engineers and data scientists to an ongoing document-format problem.

BlueFlame chose to hand much of that work to Reducto. The handoff had a concrete interface. Reducto would supply extracted content, and BlueFlame would use that output in its own GraphRAG system. Its people could work on how to retrieve information across deal documents, what the client could do through the interface and which diligence tasks an agent should perform. Bakhru wanted the engineers who understood clients’ questions to spend their time improving those parts of the product. A data scientist investigating the next spreadsheet exception was someone unavailable to improve retrieval for a whole room.

### A $5 million investment request turned into an acquisition offer

[7:17](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=437s)

Abraham returns to the independent company. BlueFlame had entered vertical AI early, when the models available to do this work were less capable. What had the founders originally hoped to build? Bakhru says they expected to be at it for a long time. They had sold their cybersecurity startup about a year in. This time they wanted to see how far they could take a company themselves.

Being second-time founders changed who they could recruit and how they could finance the work. They brought in people they had worked with before, knew one another’s families and could disagree vigorously without damaging a friendship. They could also afford to bootstrap for a while. Bakhru connects that freedom to some decisions he considers excessively cost-sensitive in retrospect. When the money for a specialist tool comes out of the founders’ own pockets, building it can look particularly attractive.

BlueFlame was about to close its first venture round. The founders had met investors on Sand Hill Road and in New York, signed a term sheet and were about a week from final signatures. There was still room in the round, so they approached Datasite about investing $5 million. They demonstrated the product at an office meeting and went to dinner that evening. Over dinner, Datasite told them it wanted to buy the whole company. The founders suddenly had a decision quite different from how to fill the remainder of a financing round.

### Datasite offered two kinds of proprietary deal data

[9:52](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=592s)

Bakhru asks Abraham to put the decision in its market context, roughly a year and a half before their conversation. Agentic coding had yet to reach the level of adoption they were now seeing. Some vertical AI competitors had already raised hundreds of millions of dollars. The founders believed that staying independent and competitive would require raising money on a similar scale, with substantial dilution. They also saw signs that the frontier model providers were developing offerings for financial services. The question was what BlueFlame could offer a finance user as those providers moved into its market.

Datasite’s rooms offered one answer. They held proprietary documents that bankers shared with buyers and advisers during a sale. Bakhru believed that joining Datasite would give BlueFlame access to a source of deal data that competing AI providers could not offer at the time. To put that advantage to use, the team had to be able to analyze a room’s documents inside Datasite’s protected environment.

A second source was Grata, a Datasite company with a database covering about 20 million private companies, including transaction and contact data. Bakhru gives the example of a private equity firm seeking add-on acquisitions for a company already in its portfolio. That is a sourcing or origination question. Grata can help someone identify companies to consider, while a data room holds the evidence buyers and advisers examine once a particular company is in a sale process. BlueFlame saw value in bringing both kinds of information into its product.

### From questions about files to diligence decks and committee memos

[11:22](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=682s)

Over the following year and a half, the companies brought their engineering teams, data and products closer together. Bakhru says the team benchmarked what happened as it brought the data closer to the models. It saw lower token costs, shorter runtimes and higher quality. He gives the direction of all three changes without specifying magnitudes or an implementation detail that would isolate the cause of each.

Demand grew along two dimensions. Bakhru estimates that Datasite’s top line grew about 40% over that period and says the volume of incoming content was growing rapidly. At the same time, customers were asking the system to do more with the documents already in a room. An individual question might require retrieving a fact. A diligence agent could iterate over documents, perform a more involved analysis and produce a PowerPoint deck or an investment committee memo. Extraction now supplied evidence for a deliverable that other people would read to make a decision.

### How Bakhru allocates 250 engineers across five initiatives

[12:44](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=764s)

Abraham asks how Bakhru now makes build-versus-buy decisions for a global workforce as model capabilities improve. Datasite spends substantially with vendors, Bakhru says, and does so deliberately. A supplier dedicated to a problem can keep working on it every day. For Datasite, assigning people to the same problem means recruiting scarce engineers and carrying the ensuing support burden alongside the product work it hired them to do.

AI coding tools have also affected whom Bakhru wants to recruit. He describes three groups. Experienced engineers who understand engineering principles and have a strong mental model of a codebase get considerable leverage from agentic coding. He has seen a more modest benefit among engineers in the middle. Junior engineers gain in a different way. Where missing documentation once left them trying to work out how the code fit together, they can ask a model to generate an explanation as they go. That observation underlies his preference for hiring strong senior engineers alongside junior people the team can develop.

He puts the allocation decision in financial terms. Using roughly a billion dollars as the order of magnitude for Datasite’s annual revenue, he estimates that the board’s ambition of 20% annual growth requires about $200 million in incremental revenue each year. The organization has about 250 engineers and five major revenue-generating initiatives. The work includes building products, consolidating existing tools and developing new commercial or monetization models. Bakhru wants as many engineers as possible on those five initiatives and as few as possible maintaining the underlying layers a vendor can operate. That is the company-wide version of his earlier decision about who should handle an unreadable spreadsheet.

### A cheaper agent model keeps calling tools

[15:54](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=954s)

Before opening questions to the room, Abraham asks where Bakhru still encounters friction. Buying parsing removed much of the work of finding the right vision model for a given chart or document. The team still has to make model choices for its agents. Bakhru describes one candidate that repeatedly issues tool calls in circumstances where another model avoids the repetition. The first is sufficiently cheaper that the team continues trying to make it behave as intended.

The comparison illustrates what Datasite has to benchmark. The price of a model call is only one input to the cost of finishing a diligence task. Repeated calls also consume time, and a fast run is of limited value if it produces a poor answer. Bakhru names three criteria: cost, speed and quality. For finance users seeking the best answer available, he gives quality first priority and works on cost and speed subject to that requirement.

### The live ingestion path from a banker’s upload to a Python query

[17:03](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1023s)

An audience member asks how Datasite and Reducto actually integrate. Bakhru says Reducto operates inside Datasite’s environment against its content. Alongside the billions of pages already on the platform, the system has to accommodate abrupt bursts. A banker can drop up to a terabyte of data into it at short notice. The annual count of rooms gives a sense of total demand, but this single-upload example explains the ingestion problem the integration has to handle.

Bakhru follows the data in order. Datasite’s ingestion pipeline takes the incoming files and passes them through Reducto. Datasite receives the extracted output and sends it to turbopuffer, its vector database. It builds GraphRAG on top of that and executes its GraphRAG queries in Python. The deployment answers two separate integration questions. The processing happens within Datasite’s environment, and Datasite retains control of the retrieval and query code that uses the parser’s output.

### VDR Bench measures what better parsing recovers

[17:46](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1066s)

A sell-side Datasite customer describes a file system containing many transactions. Their firm wants agents to use the material and wants to anonymize the material nonpublic information before passing it to a tool such as Reducto. Facing its own build-or-buy decision, the customer asks how Datasite constructed its evaluation set. Bakhru answers the evaluation question. Anonymization was the customer’s proposed requirement, and he gives no anonymization procedure.

Of Datasite’s several custom evaluations, the one most relevant to parsing is VDR Bench. The team takes a representative virtual data room containing hundreds of documents and asks a few hundred questions for which it has known answers. It reruns the evaluation after changes to almost any part of the system. Keeping the room, questions and expected answers available gives the team a way to judge a change at the level the client experiences it.

Bakhru’s benchmark for the product was what a customer could do by loading the same files into Claude Projects or OpenAI Projects and asking the same questions. He says Datasite beat that comparison and credits much of its advantage to how it extracted the content. Within Datasite’s own system, the data science team attributed a 22% improvement in recall to Reducto parsing. Bakhru gives neither the starting recall nor the scoring formula. VDR Bench was internal at the time. The team was considering releasing it and working with OpenAI on a paper.

Bakhru’s conclusion is that, for Datasite’s deal-room questions, most of the work of getting a good answer happens during extraction. The pie chart shows why. Once the parser assigns a slice’s value to the wrong category, the answering model can faithfully report the extracted figure and still give the investor the wrong answer. VDR Bench lets the team test whether a change to extraction helps it recover the right information across hundreds of questions.

## Key takeaways

- Trace a wrong answer back to the source figure and its extracted representation. In BlueFlame’s pie-chart case, the error was the association between a legend color and a slice.
- Specify both sides of a parser integration: where sensitive files can be processed and what extracted output the retrieval system will consume. Size ingestion for the bursts customers actually create.
- Evaluate agent models on completed tasks. Repeated tool calls can change runtime and cost enough to complicate a choice based on price per call.
- Test extraction changes with a fixed deal room and known-answer questions. Running the same files through a general-purpose project gives a custom system a meaningful comparison.

## Full transcript

[0:12](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=12s) I'm personally really excited about this talk. Uh, just by way of quick intro, I'm Adit, I'm co-founder and CEO at Reducto. Uh, but most importantly, we have Raj here, who is the global chief technology officer at Datasite. Um, the reason why I'm excited, uh, is I think Raj has probably one of the most interesting viewpoints for AI, in that he founded a company very early on in the vertical AI space. Um, got acquired into Datasite, and then got to see both, like, subsidiary within Datasite to thinking about Datasite as a whole. Um, I got some of Raj's stats

[0:42](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=42s) before this talk, um, where something like 60% of all transactions over $100 million happen through Datasite. Is that right?

[0:49](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=49s) That's correct, yeah.

[0:50](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=50s) Okay. Um, if we're making a leaderboard, trillion plus tokens per year, I'm sure you have plenty of other cool numbers. Um, but the goal for today is to make this conversational. We'll talk a bit about Raj's journey, um, some of the problems he's gone through, and then we'll try to open it up for Q&A as well, for whatever you all are curious about. Sound good? Cool. Um, maybe to start, do you wanna just walk through that arc and what it looked like behind the scenes?

[1:12](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=72s) Yeah, for sure. Appreciate that. Um, good to meet everyone. Raj Bakhru. Uh, my background's as a technologist, so I spent my entire career in development. Started my career at Goldman, Highbridge, Kepos Capital. Was building research and trading systems for quant funds and all sorts of fun things. Started a security firm that I sold to a private equity-backed firm. Ended up running M&A and corporate strategy for them. And then started this vertical AI solution called Blue Flame in the M&A space, um, because we sort of knew that the impact to unstructured data was gonna be pretty, pretty big.

[1:42](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=102s) Um, so we were building out for about a year, year and a half, and then Datasite approached us to acquire us. Um, and we started working with Reducto actually while we were separate. Um, so before Datasite had acquired us, um, we were a small startup, we were scrappy. Uh, we built a lot of our own parsing and doc pipeline and ingestion pipeline. Um, had a great team, a, a great CTO, co-founders. We were trying to save every dollar we could. Um, so we, when we built all of that, we were like, "This is

[2:12](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=132s) great, it's working," until it didn't.  I mean, I'll talk a little bit about that journey. Um, and then as we got brought into Datasite, so we were acquired obviously very early on. Our team was about 50 people at the time we got acquired. Um, the first thing that Datasite wanted to do with us was bring us into all of the data rooms. And if you guys aren't familiar with the data room concept, um, data rooms are basically where all of the transaction documents get stored when you're selling a company. So as Adit mentioned, something like 60% of all transactions above a, a reasonable size

[2:42](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=162s) threshold, 100 million, happen on Datasite. It's a place where people, bankers primarily, are sharing all of that data with private equity firms. Um, so it means it, there's a massive unstructured data on Datasite, right? Billions of pages of content. Every major bank is using it. Every private equity firm is using it. Every diligence provider is using it. Over 500,000 monthly active users on Datasite accessing that content. Um, and that meant that adding AI solutions was obviously top of mind for the board and everyone else who

[3:12](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=192s) was accessing the platform. Users were asking for it all, uh, day in, day out. We had a great data science team, um, from the days where my CTO, co-founder, and I sort of scrappily built the ingestion pipeline. We had obviously moved on to Reducto, recognizing that we weren't gonna have the level of quality, speed, um, recall, cost that we could get with Reducto, and I can talk about some of the numbers that led us to that, that decision, um, and why that pans out today with Data- Datasite's scale. But we

[3:42](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=222s) needed to build fast. We needed to get all of the data rooms that Datasite had, 55,000 data rooms a year, um, parsed, ingested into our vector database so that we could build out our graph RAG platform and answer questions within the data room, uh, that users were submitting at scale. Um, so that was the problem at hand. How do we, how do we build this into the VDR? They'd never done anything like this. They'd never taken their data, their content out of their Azure platform. Historically it's, you know, the walls were

[4:11](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=251s) high around the data rooms. As you can imagine, it's the world's most sensitive data. All of the most material, material non-public info out there sit in, in Datasite's virtual data rooms. Uh, so we had to figure out a way to do this securely. We had to figure out a way to do it, um, in a place that it, it, uh, uh, complied with all the regulatory compliance regimes that our clients have, and pass on to us all of the certifications, the ISO certifications and so on that, that we had. Um, so we spent a lot of time with Adit's team in figuring out

[4:41](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=281s) how to get there, how to make sure that it, all the content remained protected, um, and secure, and get us all comfortable with leveraging, uh, Reducto's tooling. The reason that we went with Reducto's tooling, as I sort of alluded to, is we basically had to, one, move fast, but two, we did a whole bunch of benchmarking actually with our data science team around sort of our scrappily built tooling for extraction and parsing and ingestion. And what I'll say about sort of that landscape is that, um, it's all edge cases. Like,

[5:11](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=311s) literally as, everything is an edge case. I remember getting... You know, after we had built this, we were working with some of the world's largest private equity firms, and they would send to us, um, examples where our RAG platform got a question wrong. And actually, Carlos here will, will remember some of these.  Um, you'll dive into it and be like, "Well, it got it wrong because the shade of blue on the legend of this pie chart didn't correlate perfe- perfectly, and the LLM just read it wrong, even though we used the best vision model in the world." And actually, if you were to pass that th- same thing as a

[5:41](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=341s) pie chart or that pie chart, um, as a bar chart to that same model, it would've gotten it right. It just, the devil's in the details on all of this stuff. Um, on Excel sheets, we had to deal with merge cells, and as you guys know, finance spreadsheets are the most complicated documents in the world. Like, it is hard for a human to understand these things. Um, so getting the models to actually understand the way to parse this, this content out and understand the content was very, very difficult. Um, so we found that we were just spending an enormous amount of support time dealing with all

[6:11](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=371s) these edge cases. And if I sort of think about the buy verse build trade-off, what do I want my limited set of high-quality engineers working on? It is not the support cases of making sure that we're extraction-- extracting content properly, um, when that's already been solved. Like, you know, rebuilding the wheel doesn't make sense. Um, so we, we got a tremendous amount of benefit and leverage out of saying, "Let's, let's pass that on to Reducto. Let's let our engineers work on value add revenue generating business problems,

[6:41](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=401s) focus on building agents that our clients care about, executing on the interface and the RAG platform." Uh, building out GraphRAG was actually a big thing for us, and we leveraged the outputs of Reducto in that process. So we wanted our data science team focused on that, not on how do I get the right data out of a spreadsheet, um, and dealing with all the pains that we experienced on that. Then you, you run into the pains of, like, if it's an XSLB file, um, macro-enabled, if it's, uh, a template file, it just, it's endless.

[7:11](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=431s) Literally, the end-- edge cases are endless, so we're very thankful to Reducto for solving a lot of those problems for us.

[7:17](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=437s) I appreciate that. Um, at this point, we've had, like, a year of context, but I, I think one of the most interesting things about your journey is that you were very early to the vertical AI app space. Um, and the problems that you work on are hard, um, for many of the reasons that you mentioned, but also, like, things beyond that. Um, and especially earlier with the model landscape as it existed, I'm sure it felt even more insurmountable. Um, so I'd love to start maybe just by hearing a bit more of the pre-Datasite journey of, like, how you got started with Blue Flame, what you

[7:47](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=467s) were hoping to achieve, um, and how joining Datasite has kind of changed that.

[7:52](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=472s) Yeah. Um, so when, when we launched our, our startup, startup, we had sort of-- we'd sold a startup very early on before. The last startup, cybersecurity startup, we sold a year in, and we started this one saying, "We are not going to sell quickly. We are going to see this out, and it's going to be a long time, and we're going to be a unicorn, and do all those fun things that we wanted to do." Um, so that was, that was the initial state. Brought in a lot of folks that we had worked in-- worked with before, and that's one of the nice things of being a second-time, uh, founder is you kind of know the

[8:21](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=501s) network, you know the people you want to bring in. Uh, so we were very fortunate to bring in a lot of those people and start with a lot of the people that we knew and trusted, um, and we could yell at each other and still be really good friends at the end of the day. Um, and we knew each other's families and all that. So that really nice experience to start as a second time. Uh, we also had the ability, thankfully, to fund ourselves to a large extent this, this second time. Uh, so we bootstrapped for a while and started building. Now, that, again, made us do some things that might have been a little excessively cost sensitive, um, and not the most optimal decisions,

[8:51](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=531s) but that's what you do as a startup, um, and when it's your own money, especially. So we had a really great experience building that out. Uh, we were actually going to do our first VC round. We had our LO- LOI signed, term sheet signed. Um, we had done sort of the dating process with all the VCs out there, the Sand Hill Road walk and all the New York ones. Um, and we were, you know, a week away from sort of final, final signatures, and there was additional capacity on our round. Um, and they said, "You should go, you know, see if these

[9:21](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=561s) folks are interested." And so we, we approached Datasite and we're like, "Do you guys want to put in five million in this round?" And we demoed and showed them a bit of what we were doing, and they said to us, "Actually, we just want to buy the whole company." Um, we're like, "Okay." It literally happened over dinner. Um, like, whatever, meeting in the office, subsequent dinner that night, and then they told us that. Um, so we, we very quickly sort of figured it out. And the reason that we thought it was really compelling, um, was if-- You know, this was about a year and a half ago now. At that point, sort

[9:52](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=592s) of Claude and OfficeSuite hadn't taken off. You sort of have to go back to remember where we were a year and a half ago. It feels like forever ago, and the world was entirely different. Um, agentic coding hadn't taken off to the level it has now. Um, so we knew that we had to raise a massive amount of money. Our competitors in the vertical AI space, um, some of whom work with Reducto, had raised hundreds of millions of dollars, literally. Um, so we kind of knew that if we're going to remain competitive, we were going to be raising hundreds of millions of dollars. We would be taking masses-- mass dilution in doing that. Um,

[10:22](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=622s) we also knew that staying competitive against the frontier models was going to be increasingly difficult, particularly as they were starting to launch vertical strategies. And at that point, sort of OpenAI for FS and Claude for FS hadn't been fully announced, but there were some rumors around it. They'd made some, some ins-insinuations about it. Um, and we knew that just to remain competitive and differentiated, that we would have to bring in new types of data. And to us, the BDRs were this incredible, valuable resource of prop data that

[10:52](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=652s) had never been opened up to anyone in the world, and we would be the only ones with access to it at that time. So that was super compelling. Um, Datasite also owns a private company database firm called Grata. Has about twenty million private companies in its database, has transaction data, contacts, et cetera. Used very heavily in private equity and banking worlds for, um, sourcing and business development and origination to help me find add-ons for my platform portfolio company. Uh, very classic Grata type query. That data would be brought

[11:22](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=682s) into Blue Flame as well. So now we had these two really nice proprietary data sets feeding us in a vertical AI space where no one else would have that. Really nice and differentiated, and that's played out really nicely. So over the, over the course of the, uh, last year and a half, we've been bringing those really tightly together and, um, that's, that's been a really nice story for us to combine engineering teams, combine data, combine solutions. Um, and we actually benchmarked that, like the, the ability or the, the fact that we brought the data closer to the models- Saves on token costs,

[11:52](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=712s) it saves on run, run times, um, and increases quality of what we're doing. So that's been a... That's been, um, quite a nice thesis to see pan out. Um, all of that data obviously requires scale, so that's, that's again where Reducto comes in and helps us with a lot of that ingestion processing and, and so on. Datasite over the course of the last year and a half since we've joined them has grown, I wanna say about 40% top line.

[12:18](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=738s) Mm.

[12:18](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=738s) Um, so just the amount of content that's coming into the platform is, is growing pretty quickly and pretty massively. Um, the demand on our AI solutions have gone up, so now people aren't just saying, "I wanna ask questions against documents." The agents are obviously iterating against those documents, doing much more fanciful diligence analyses. They're generating PowerPoint decks, IC memos, and so on. Um, so it's a pretty fulsome suite of what they can, what they can accomplish with the solution together.

[12:44](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=764s) Yeah. Um, it's an interesting position to be in, in that I, I think you have one of the most compelling data sets in the industry. Um, you're in a position now where obviously it's a global workforce. We were talking about all the different offices and all of that, um, and model capabilities have gotten a lot better. Um, so more broadly, like, how do you think about what you invest in internally versus what you build, uh, and buy externally as well? Um, and also just like what sorts of things are top of mind as you think about the coming year?

[13:14](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=794s) Yeah. Um, great question. So we think a lot about the build versus buy, right? We, we work with a lot of vendors. We spend a lot with a lot of vendors.  Um, and that's an intentional decision. Um, if I think about what I want my, my most talented people working on, it's not things that vendors can easily handle and are dedicated to handling and solving every day. Um, it's really hard to find really high quality, good engineers, right? The fact of the matter is where everyone in market is trying to hire the top tier talent of engineering

[13:44](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=824s) talent right now because what we're seeing, what you guys are all probably seeing i- is with AI coding, the folks who get the most leverage and value out of AI coding are the folks who have really strong mental models of code bases, really understand engineering principles, have done it for a long time. Um, so they've all been 10X'd, uh, you know, over their normal 10X, uh, with AI... with agentic coding. The middle layer, so-so. Um, the bottom layer is actually getting a lot of leverage, and now they don't, they don't sort of, uh,

[14:14](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=854s) have the burden of trying to figure out documentation that doesn't exist and all of that. The LMS can build that dynamically for them. Um, so we're seeing sort of a barbell effect where we wanna hire really, really strong talent and really junior talent and build that up. Um, where I want those people working is obviously not on support cases. It's not on things that are already solved. Um, it's on the hardest and most pressing problems for the business. At our scale, so Datasite's not a billion dollar top line, but general range of

[14:44](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=884s) a billion dollar top line. Um, at our scale to grow 20% a year, which is our ambition, our board's ambition, uh, we need to be generating two hundred million dollars of revenue, incremental revenue every year. And if you ask yourself how you're gonna go do that, um, it's gonna be by putting new AI products out in market, leveraging, getting on top of token flow, expanding solution set that you're delivering to your clients. Um, reinventing the wheel is certainly gonna get in the way of that. Um, having internal teams

[15:14](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=914s) trying to, uh, iterate on some of the existing solutions that are, that are already out there and deal with the support burden of those is not the best use of my talent. I have about two hundred and fifty engineers in the org. So if I allocate out where I want those two hundred and fifty focused, we have five major initiatives across the firm. Um, those are five revenue-generating projects that we're putting out, new, new builds or consolidations of tools, new commercial or monetization models. Um, I want those two hundred and fifty pe- people split across those five, and

[15:44](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=944s) I want as few working on the layers underneath that as possible. And if I can outsource everything on the under-- the layers underneath that, that's an ideal, um, ideal split of time for my, my team.

[15:54](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=954s) Yeah. Uh, before I turn it over for questions, I realize we have two minutes. I guess maybe one last thing is like where do you find friction today? Um, what are problems that you would hope people in the audience would solve for you?

[16:05](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=965s) Um, we still struggle with getting the models to do exactly what we want all the time.

[16:11](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=971s) Mm.

[16:11](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=971s) Um, we do a lot of benchmarking of all of the models, and actually, you know, if you go back to pre-Reducto world, we wasted a lot of time on trying to figure out the right model for the right type of chart, for the right type of, uh, document that we're passing in. So we've, we've gotten rid of a lot of that, but we still deal with it on the agent front-

[16:27](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=987s) Mm.

[16:27](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=987s) -trying to benchmark out, you know, why is six-one sole tool spam, doing tool call spamming when Opus five-five is not? But then six-one sole is so much cheaper that we wanna move in that direction, and we're trying to get it to go, to go there. Um, so we do a lot of sort of the optimization of the three factors we care about, which is cost, speed, and quality. Quality is top, top of mind for everyone in our space, right? Like any finance user will pay any amount of money to have the best answer. Um, so we need to make sure we're

[16:57](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1017s) always optimizing on quality while trying to solve the other two.

[17:00](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1020s) Yeah. Makes sense. Any questions from the room? Yeah.

[17:03](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1023s) Can you zoom a little bit about the integration between Datasite and Reducto?

[17:10](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1030s) Yeah. So we have an environment where Reducto's, um, operating in our environment against all of our content. We have billi- billions of pages of content. We have very big spurts of content when bankers are dropping in large data sets, potentially up to a terabyte of data can get dropped in on a moment's notice. Uh, so we have an ingestion pipeline that takes all of that, passes it through Reducto. We get that back. We, we leverage TurboPuffer as our vector database, so we're passing it off to TurboPuffer. Um, and then we build our graph rag on top of that. We're executing all of our graph rag queries just,

[17:40](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1060s) uh, using pro-- Python code against all of that.

[17:45](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1065s) Yeah.

[17:46](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1066s) Um, I'm on the sell side of things. Um, we're-- we are actually customers-

[17:52](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1072s) Yeah.

[17:52](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1072s) -of Datasite.

[17:52](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1072s) Every sell side is.

[17:54](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1074s) Um, curious about Reducto. So we're sitting on a transaction. Like, we have a, a file system with a ton of transactions, and we wanna make that available to agents, but there's a way of... Like, obviously it's MNPI. We wanna anonymize that before it makes its way into something like Reducto.

[18:13](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1093s) Mm.

[18:13](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1093s) We're gonna either build this or buy this, so it's very helpful conversation. That being said, giving it to the agents, you s- you mentioned before that you built a custom eval, and so I, I'm curious about how y'all went about this eval-

[18:30](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1110s) Yeah.

[18:30](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1110s) -which is potentially open sourced.  Um, but, like, curious about how you guys-

[18:36](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1116s) Yeah.

[18:36](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1116s) -went about that eval set.

[18:37](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1117s) So our, our custom eval that's related to Reduc-- We have a few custom evals, but the one that's most relevant to Reducto is what we call VDR Bench. We haven't open sourced it. We're actually thinking about doing that and thinking about partnering with OpenAI and a paper around it. Um, what we're doing with VDR Bench is we basically take a common virtual data room of hundreds of documents, um, and we ask a few hundred questions against that data room with a known set of answers against those. Um, we run that every time we make changes to, to pretty much anything. Um, the scores on

[19:07](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1147s) that we expect to be higher than any of the, uh, frontier labs directly. So if you were to, for example, drop that same set of files into cloud projects or OpenAI projects, we need to make sure that we're beating that. Um, and a lot of the reason that we beat that and our data science team benchmarks sort of our recall rates at, you know, improving twenty-two percent because of our Reducto parsing, um, you know, a lot of that's because of the, the way we pull out the content. It's actually not the other side of it. A lot of people think that sort of all the work needs to be done on the other side, the answering side, and most of the

[19:37](https://www.youtube.com/watch?v=0PqGT2L6cZM&t=1177s) work actually is on the extraction side. You get that right, and everything else can follow.
