Anthropic Engineer Builds System to Keep Claude Reliable

Anthropic has tasked IEEE Senior Member Gurmeet Saran, who joined the company’s technical staff in 2025, with developing the data infrastructure needed to observe, audit, and verify the internal logic of its Claude model family. As large language models transition into autonomous agents taking sequences of actions, traditional software observability tools fall short. Saran’s work focuses on capturing model logic, ensuring telemetry accuracy, and enforcing strict privacy safeguards so sensitive data remains protected as it flows through the underlying architecture.

Saran brings over two decades of enterprise data systems experience to Anthropic’s San Francisco headquarters. After earning a bachelor's degree from the University of Kalyani in 2004 and a master's degree from the Birla Institute of Technology and Science in 2010, he began solving data inconsistency issues at Yahoo in 2007 by creating the standard definitions behind Yahoo! Digits. Later at Intuit, he helped engineer real-time fraud detection systems for TurboTax and QuickBooks that complied with Internal Revenue Service privacy regulations, cementing his philosophy that governance must be designed into systems from inception. At Meta starting in 2016, Saran built Metric360 to resolve systemic errors in video view metrics, creating a unified metric framework used across the platform until his departure in 2021 before serving as principal software engineer at Wealthfront.

For machine learning engineers and developers building on Claude, Saran's efforts address a critical bottleneck in agentic AI deployment: explainability and trust. Observability in LLMs requires determining exact context and decision paths while protecting end-user privacy. By embedding strict governance, standardized metrics, and auditability directly into Anthropic's data pipeline, the initiative aims to give enterprise teams predictable, verifiably accurate outputs when running Claude in high-stakes operational environments.