Why Datasite Bought Its Document Parser Instead of Building It

Adit Abraham, co-founder and CEO of Reducto, introduces Datasite global CTO Raj Bakhru. Bakhru has seen vertical AI from both sides: he founded BlueFlame, then joined Datasite when it acquired the company.

Abraham says about 60% of transactions above $100 million run through Datasite; Bakhru confirms the figure. Datasite’s virtual data rooms hold the documents exchanged among bankers, private equity firms and diligence providers during transactions.

Bakhru began by building research and trading systems for quantitative funds, then founded a security company. He later started BlueFlame to apply vertical AI to mergers and acquisitions, anticipating the importance of unstructured data.

As a small startup, BlueFlame built much of its own parsing and document-ingestion pipeline to save money. The system worked until the diversity of customer documents exposed a growing number of failures and support cases.

After roughly a year and a half, Datasite acquired BlueFlame. Its first priority was to bring the technology into its data rooms. Bakhru says the platform handles billions of pages, more than 500,000 monthly active users and about 55,000 data rooms a year.

The task was to parse and ingest that content into a vector database, build a graph RAG platform and answer users’ questions at scale. Datasite had historically kept its sensitive transaction data within Azure, so the team had to address security and regulatory requirements before using Reducto.

Bakhru says the team benchmarked its own extraction and ingestion tools. It chose Reducto for a combination of quality, speed, recall and cost, as well as the need to move quickly.

Parsing turned out to be a succession of edge cases. One model misread the relationship between a pie-chart legend and its colors, although the same information in a bar chart worked. Financial spreadsheets introduced their own difficulty, including merged cells and complicated layouts.

The build-versus-buy decision became a question of where scarce engineering time should go. Bakhru preferred to pass extraction problems to a specialist and have his engineers work on customer-facing agents, the interface and the RAG platform.

His data science team could then focus on graph RAG rather than continually repairing spreadsheet extraction. Binary Excel files, macros and templates were among the formats that kept generating new cases.

Bakhru describes BlueFlame’s early days: after selling an earlier company quickly, the founders intended to build a long-term standalone business. They brought in people they knew and trusted and initially funded much of the work themselves.

BlueFlame was close to closing its first venture round when it approached Datasite about a possible $5 million investment. After a demonstration and dinner, Datasite instead proposed buying the company.

Bakhru says the team expected that staying competitive in vertical AI would require raising hundreds of millions of dollars, with substantial dilution. Frontier model providers were also beginning to pursue industry-specific strategies.

Datasite offered access to distinctive data: virtual data rooms and Grata, a private-company database that Bakhru says covers about 20 million companies. For BlueFlame, combining those data sources with its AI product was a compelling reason to join.

He says bringing data closer to models improved quality and reduced token costs and runtime in the team’s benchmarks. Reducto’s ingestion and processing became important as the volume of incoming content continued to grow.

Bakhru estimates that Datasite’s revenue grew about 40% over the previous year and a half. Customer requests have expanded from asking questions about documents to agents that conduct diligence and generate presentations and investment committee memos.

Asked what to build or buy, Bakhru says Datasite intentionally uses many vendors. He wants its strongest engineers focused on difficult business problems rather than components that specialist suppliers already maintain.

He argues that experienced engineers with strong mental models benefit greatly from AI coding tools. He also sees junior engineers gaining leverage, leading him to describe a hiring preference for both highly experienced and junior talent.

Bakhru says the company’s ambition is to grow around 20% a year at a business scale near $1 billion in revenue. In his calculation, that means finding roughly $200 million in additional annual revenue through new AI products and expanded offerings.

He has about 250 engineers and five major initiatives, including new products, consolidation and commercial models. His preferred allocation puts them on those revenue-generating projects, with as few people as possible maintaining underlying layers that can be outsourced.

The remaining challenge is making models do exactly what the application needs. Teams still compare model behavior, tool-call patterns, speed, cost and quality. Bakhru says quality is the leading concern for finance users, while the other two factors must also be managed.

In the audience Q&A, Bakhru says Reducto operates in Datasite’s environment. Large transaction uploads can arrive in bursts of up to a terabyte; the pipeline processes the content, passes it to a vector database and builds graph RAG on top.

An audience member asks about evaluation for agents working with sensitive transaction files. Bakhru describes VDR Bench: a virtual data room containing hundreds of documents, with hundreds of questions and known answers. The team reruns it after changes.

The benchmark is internal, though Bakhru says the team is considering publishing it. It aims to beat using the same files directly with frontier-model projects. He reports a 22% improvement in recall from Reducto parsing and argues that extraction quality strongly affects everything that follows.