Datasite global CTO Raj Bakhru joins Reducto co-founder and CEO Adit Abraham for a 20-minute AI Engineer conversation about document ingestion and engineering priorities.

Parsing failures become answer failures

  • Bakhru’s startup BlueFlame built its own ingestion pipeline before switching to Reducto, then being acquired by Datasite.
  • A vision model misread the colors in a pie-chart legend; the same information in a bar chart could produce the correct answer.
  • Merged spreadsheet cells, binary workbooks, macros and templates created a long tail of support work.
  • He wanted engineers building customer workflows and retrieval rather than repeatedly repairing extraction edge cases.

The deployment has to fit the data

  • Datasite hosts virtual data rooms: collections of confidential documents shared during company transactions.
  • Bakhru describes billions of pages, roughly 55,000 data rooms annually, and uploads that can arrive in bursts approaching a terabyte.
  • His integration account places Reducto inside Datasite’s environment, feeding parsed content into turbopuffer, a vector database.
  • Proprietary Python code runs graph-based retrieval and generation over that content, combining document search with relationships between information.
  • Security and client compliance requirements shaped the integration before the team could expand access to the documents.

Keep scarce engineering time on the product

  • Bakhru describes about 250 engineers and five major revenue-generating initiatives competing for their attention.
  • Buying a specialist layer can free that team to improve analysis, interfaces and customer-facing agents.
  • Datasite’s confidential documents and Grata’s private-company data form the differentiation he wants those products to use.
  • He also describes a hiring preference for strong senior engineers and juniors, based on his experience with AI coding tools.

Test the whole question-answering system

  • Their private VDR Bench uses hundreds of documents, a few hundred questions and known answers, and runs after changes across the system.
  • The comparison they target is the same files supplied directly to frontier-model project interfaces.
  • Bakhru reports a 22% recall improvement from Reducto parsing; he gives no public protocol or clarity on relative versus percentage-point improvement.
  • Model selection remains a balance of quality, speed and cost, including whether a cheaper model makes excessive tool calls.

“Most of the work actually is on the extraction side.” — Raj Bakhru, 19:35

Caveat: This is a supplier-customer conversation. The internal benchmark and performance claims are described by Datasite, with no public results for independent comparison.