Atul Ramachandran is the co-founder and chief technology officer of Filed, which builds AI systems that help tax professionals prepare returns while retaining control over consequential decisions. He also created NodeGui and React NodeGui, open-source frameworks for building native desktop applications with JavaScript and React.
His earlier engineering career included McKinsey & Company, payments company iZettle, and Swedish financial-technology company Anyfin, where he advanced into technical leadership. His technical writing covers Docker, GraphQL authorization, React renderers, and Node.js extensions. In 2019, he introduced NodeGui and React NodeGui, developing an approach to cross-platform desktop software that combines familiar web-development tools with native interfaces.
Ramachandran later served as chief technology officer at Magic Returns before founding Filed with Leroy Kerry. Filed raised $17.2 million and joined the 2025 AICPA and CPA.com startup accelerator, building software that extracts information from client documents, adapts to accounting firms’ practices, and integrates with professional tax systems.
Designing AI that professionals can actually delegate to
- Agentic delegation: Chat interfaces keep professionals waiting for responses, while citations can shift verification work back onto them. Ramachandran’s alternative gives background agents substantial assignments and casts professionals as supervisors who intervene when judgment is needed.
- Automatically captured agent skills: Agents should learn a firm’s conventions and institutional knowledge from ordinary product usage, without requiring customers to maintain separate instruction libraries.
- Human-controlled agent workflows: Task lists, execution traces, and inspectable generated values make long-running work legible. Agents should pause when assumptions require professional judgment, and potentially destructive changes to tax software should require advance approval.
- End-to-end tax workflow reliability: Filed combines foundation models with specialized agents, document processing, validation, and existing-software integrations. Its tax-accuracy comparison positions these safeguards as essential to professional work that still requires human review.
- Weekly active sessions: Ramachandran measures completed human or agent tasks instead of treating time spent inside an application as the primary signal of value. Effective delegation should increase completed work while reducing the attention customers must supply.
