Teamday
← Back

Datasite CTO: Why We Stopped Building Our Own Parsing

AI Engineer
agentsenterprisebusinessraginterview

How Datasite Brought AI Agents Into the World's M&A Data Rooms

Raj Bakhru is the global CTO of Datasite, the virtual data room platform where, by his count, about 60% of all transactions over $100 million happen. He co-founded Blueflame AI, a vertical AI startup for M&A, which Datasite acquired in July 2025. In this AI Engineer conversation with Adit Abraham, co-founder and CEO of Reducto, he explains why a company processing more than a trillion tokens a year chose to buy document parsing instead of building it.

Blueflame built its own parsing, until it broke: As a bootstrapped startup, Blueflame wrote its own parsing and ingestion pipeline to save money. It worked until private equity clients sent back wrong answers. One failure came from a pie chart whose legend colors didn't quite match, which the best vision model misread, though the same data as a bar chart came out right. Finance spreadsheets brought merged cells, macro-enabled files and templates. "It's all edge cases. Like, literally, everything is an edge case." The support time spent on those cases is what pushed Blueflame to Reducto before the acquisition.

The scale of Datasite's data: Datasite rooms hold billions of pages of the most sensitive material non-public information in finance, used by every major bank and private equity firm, with over 500,000 monthly active users. The first job after the acquisition was to parse and ingest 55,000 data rooms a year into a vector database for a GraphRAG platform. Datasite had never moved that content out of its Azure environment, so Reducto runs inside Datasite's environment, under its regulatory and ISO commitments. Bankers can drop up to a terabyte into a room at short notice. Parsed output goes to turbopuffer as the vector database, and Datasite runs its GraphRAG queries on top in Python.

Extraction does most of the work: Datasite's custom eval, VDR bench, asks a few hundred questions with known answers against a typical data room of hundreds of documents, and runs on almost every change. The bar is to beat the same files dropped into Claude or OpenAI projects. Bakhru credits Reducto parsing with a 22% recall improvement and says people look in the wrong place: "most of the work actually is on the extraction side. You get that right and everything else can follow."

Build versus buy at $1 billion revenue: Datasite's top line is in the range of $1 billion, and Bakhru says it has grown about 40% since the acquisition. Growing 20% a year means "we need to be generating $200 million of revenue, incremental revenue every year." He has about 250 engineers split across five revenue-generating initiatives and wants as few of them as possible on the layers underneath. Talent drives the same answer: AI coding gives the most leverage to senior engineers with strong mental models of codebases, and also lifts juniors, so Datasite is hiring at both ends of a barbell.

Proprietary data is the moat against frontier labs: Blueflame sold rather than raise hundreds of millions to compete with funded vertical AI rivals and frontier labs moving into financial services. The draw was data nobody else had: Datasite's data rooms and Grata, Datasite's private company database of about 20 million companies, used for sourcing add-ons and origination. Bakhru says they benchmarked that bringing the data closer to the models cut token cost and run time and raised quality. Agents now go beyond question answering to diligence analyses, PowerPoint decks and IC memos.

5 Lessons from Datasite on Buying AI Infrastructure

  • Parsing is endless edge cases - Charts, merged cells and macro files are where RAG answers fail, and fixing them by hand eats engineering time.
  • Measure against the frontier labs - VDR bench requires Datasite's answers to beat the same documents in Claude or OpenAI projects.
  • Put engineers on revenue - With 250 engineers and a $200 million yearly growth target, Bakhru staffs the five revenue initiatives and buys the layers underneath.
  • Run vendors inside your walls - For MNPI, the parsing vendor runs in Datasite's own environment rather than receiving the data.
  • Quality comes first in finance - Datasite optimizes cost, speed and quality, but "any finance user will pay any amount of money to have the best answer."

What Datasite's Choice Means for Enterprise AI Teams

Bakhru's rule is to ask what your scarcest engineers should work on and buy everything else that a dedicated vendor solves every day. For data-heavy agents, the payoff sits upstream: clean extraction lifts recall more than tuning the answering model. For another finance team's view of when to build in house, see Ramp Labs on post-training a specialist model.