Why Datasite Stopped Building Its Own Parsing

Adit Abraham, Reducto, Raj Bakhru, Datasite19:49 · Nov 2026 · 3,645 views
Thumbnail for Why Datasite Stopped Building Its Own Parsing Watch on YouTube
TL;DR
  1. 1

    Datasite moved from in-house document parsing to Reducto because financial documents produced too many difficult edge cases for its engineers to maintain.

  2. 2

    Raj Bakhru wants Datasite's 250 engineers focused on new AI products and revenue-generating problems instead of rebuilding infrastructure that specialist vendors already handle.

  3. 3

    Datasite's VDR Bench evaluates answers over hundreds of documents, and better parsing raised recall by 22% in the company's internal benchmarks.

Summary

Raj Bakhru describes how his BlueFlame startup became part of Datasite and brought AI into a platform containing billions of pages from sensitive deal documents. BlueFlame initially built its own extraction and ingestion pipeline, but financial files exposed endless problems with charts, merged spreadsheet cells, macro-enabled files, and templates. Datasite decided that Reducto could provide better quality, speed, recall, and cost while freeing its engineers to build agents, graph retrieval, and customer-facing products. Bakhru also explains why the acquisition gave BlueFlame access to proprietary transaction data and Datasite's private-company database. He wants the company's 250 engineers concentrated on five major initiatives rather than support work for solved infrastructure problems. The remaining challenges involve choosing models and balancing cost, speed, and answer quality. Datasite tests changes with VDR Bench, a private evaluation over hundreds of documents and questions.

Key ideas
01:18

BlueFlame began as a vertical AI company for M&A data

Raj Bakhru started his career building research and trading systems at Goldman, Highbridge, and Kepos Capital before founding BlueFlame. The company focused on M&A because he expected unstructured data to have a large impact there. Before the acquisition, the small team built its own parsing, document, and ingestion pipelines to save money. That approach worked for a while, but its limits became clear as the company handled more demanding documents.

02:21

Datasite gave BlueFlame access to a large and sensitive document base

After acquiring BlueFlame, Datasite wanted the team to bring AI into its virtual data rooms, where transaction documents are stored during a company sale. Bakhru describes billions of pages, more than 500,000 monthly active users, and about 55,000 data rooms processed each year. The team had to move content out of Azure while keeping it protected and meeting the certifications and regulatory requirements of Datasite's customers.

04:54

Document parsing failed because every file contained different traps

Bakhru says parsing financial documents is 'all edge cases.' A pie chart could be misread because the shade of blue in its legend did not match perfectly, even with a leading vision model. The same model could answer correctly when the chart was changed to a bar chart. Spreadsheets added merged cells, macro-enabled XLSB files, and templates. The team spent too much support time handling these cases instead of building product features.

06:12

Datasite chose Reducto so its engineers could work on customer value

The build-versus-buy decision came down to where scarce engineering time would produce the most value. Datasite found that Reducto offered better quality, speed, recall, and cost than its own parsing system. Bakhru wanted engineers working on agents, the user interface, the retrieval platform, and graph retrieval. He did not want them maintaining extraction behavior that a specialist vendor was already addressing.

09:40

The acquisition combined proprietary data with Datasite's distribution

BlueFlame had planned to raise venture capital, but Datasite offered to buy the company instead. Bakhru says remaining independent would have required raising hundreds of millions of dollars against competitors and frontier model providers. Joining Datasite gave the team access to virtual data room content and Grata, Datasite's database of about 20 million private companies, along with transaction and contact data. Combining those datasets with the engineering teams became part of the product strategy.

11:33

Putting data closer to models lowers several operating costs

Bakhru says bringing Datasite's proprietary data closer to the models saves token costs, reduces run times, and improves quality. The combined system now supports questions over documents as well as more involved diligence analysis and generation of PowerPoint decks and investment committee memos. He says the amount of content entering Datasite has grown quickly, increasing demand for these AI features and for scalable ingestion.

13:13

AI coding is changing which engineers Datasite hires

Bakhru describes a barbell effect in engineering hiring. Experienced engineers with strong mental models of codebases and engineering principles get substantial value from agentic coding. Junior engineers also gain help with documentation and other tasks, while the middle layer sees less benefit. Datasite therefore wants strong senior talent and strong junior talent, with both groups focused on difficult business problems rather than support cases.

17:02

VDR Bench tests the whole system, starting with extraction

Datasite's VDR Bench uses a virtual data room containing hundreds of documents and asks hundreds of questions with known answers. The company runs it whenever it changes a significant part of the system and compares its scores with frontier model products. Bakhru says Datasite's recall improved by 22% because of Reducto's parsing. His point is that answer quality often depends first on extracting the source content correctly.

"And what I'll say about sort of that landscape is that it's all edge cases. Like literally everything is an edge case."Raj Bakhru04:59
Who should watch
  • You are deciding whether to build document extraction and ingestion internally or buy it from a specialist vendor.
  • Your team is building AI over sensitive financial or transaction documents and needs to think about security, compliance, and evaluation together.
  • You manage an engineering organization and want to allocate scarce senior talent between infrastructure maintenance and new AI products.