▶ Watch ↗19:48
Conference affiliation: Datasite
Raj Bakhru is chief technology officer of Datasite and co-founder of Blueflame AI, which builds search and automation software for investment professionals. His work has moved from quantitative trading systems through cybersecurity to AI for dealmaking. Across those fields, he has built software around the information and workflows investment firms use to make decisions.
Raj studied computer engineering at Columbia University and began his career developing research and cross-asset trading software at Goldman Sachs Asset Management. At Highbridge Capital, he led the team building a proprietary order and execution management system. He subsequently joined the founding team at Kepos Capital and led firm-wide software development.
In 2014, Raj and Marc Lotti co-founded Aponix, a cybersecurity and technology risk business serving the investment industry. ACA acquired Aponix in 2015, making it the foundation of its cybersecurity practice. Raj later served as ACA’s chief innovation officer and chief strategy officer. His responsibilities expanded from developing software to corporate development, acquisitions and the ComplianceAlpha regulatory technology platform.
Raj and Henry Lindemann founded Blueflame AI in 2023. Blueflame connected internal records, proprietary research and public information so investment professionals could search and analyze material scattered across different systems. Its applications covered deal sourcing, diligence, market research and fundraising.
In his first-anniversary account, Raj described complementary ways to turn that information into work. Nexus organized unstructured information into a searchable database. DDQ Assistant helped compile responses to due-diligence questionnaires and requests for proposals. Blueprints let users define multistep AI workflows through natural-language instructions. These products extended the company’s search capabilities into recurring tasks that required gathering and using information from several sources.
Datasite acquired Blueflame in July 2025, when Blueflame had about 50 employees. Raj had initially intended to keep his second startup independent after selling his cybersecurity business early. In his AI Engineer NYC discussion, he explained why that position changed: competing with heavily funded startups and frontier model providers would require substantial capital, while Datasite offered access to transaction documents and Grata’s private-company intelligence. The acquisition brought Blueflame’s software together with information used both to find companies and to investigate them during a deal.
Raj became Datasite’s CTO in September 2026, extending his responsibility from Blueflame to the wider technology organization. At AI Engineer NYC, he described leading about 250 engineers. An early integration task was bringing Blueflame’s retrieval capabilities into Datasite’s virtual data rooms, where bankers share confidential transaction documents with prospective buyers.
Raj’s essay on buying versus building AI distinguishes tools that create an investment firm’s advantage from software that supports general productivity. Proprietary data, models or requirements can justify custom development; vendor capabilities, maintenance, integration and security costs also belong in the decision.
Blueflame’s document parser gave him a concrete reason to revisit an early build decision. The team initially developed its own ingestion pipeline to conserve cash. Then difficult charts and financial spreadsheets began consuming engineering and support time. In one case, a wrong answer traced back to a vision model matching a pie-chart slice to the wrong shade of blue in its legend. Merged spreadsheet cells created another class of interpretation problems.
Blueflame adopted Reducto before the acquisition so its engineers could concentrate on retrieval, interfaces and agents for deal teams. At Datasite, the team used the parsed output to build graph-based retrieval across virtual data rooms. Raj reported a 22% improvement in recall from Reducto parsing on Datasite’s internal VDR Bench: hundreds of documents paired with a few hundred questions and known answers, rerun when the system changed. The result was an internal benchmark claim, but it illustrated his technical position: improving extraction could improve answers without concentrating all the work on the model’s final response.
Two other priorities shape how Raj applies those capabilities:
▶ Watch ↗19:48
Loading more talks and highlights…