Meet the Anthropic Engineer Keeping Its AI Agent Trustworthy
IEEE Spectrum Amanda Davis
An Anthropic engineer is trying to make Claude easier to audit and trust. The hard part isn’t just answers — it’s proving what the AI used, and keeping private data private.
Based on reporting by IEEE Spectrum, Amanda Davis — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Gurmeet Saran spends his days on a problem most users never think about: when Claude answers fast, how do you know where the answer came from, and whether the system can be trusted to show its work? At Anthropic in San Francisco, he is building infrastructure to understand the mechanics behind the chatbot’s behavior and to make sure the architecture produces reliable outputs.
His fix-and-test instincts were shaped long before AI. As a student in Kolkata, he was fascinated by the simple fact that a computer would do exactly what it was told. Later, after earning degrees in electronics and telecommunication engineering at the University of Kalyani and in software engineering at the Birla Institute of Technology and Science in Pilani, he started building data systems at Yahoo in 2007.
Yahoo is where he ran into a classic problem with a modern smell: different teams were calculating the same website metric in slightly different ways and getting different answers. Saran helped create a central data repository with defined metrics, including a model for what counted as a daily active site user. That work turned into Yahoo! Digits, an internal analytics tool used across the company.
He carried the same idea to Intuit, where he helped build analytics infrastructure for products like TurboTax and QuickBooks, and where privacy and governance had to fit the system from the start because of real-time fraud detection and U.S. Internal Revenue Service rules. At Meta, he built Metric360 after the company acknowledged it had overstated video views. The point, again, was a single source of truth.
Now the stakes are different, and messier. Saran says AI agents are hard to observe because they take sequences of actions, pull from different information, and have to protect sensitive data while doing it. He compares the job to making sure medical instruments are calibrated correctly before a doctor makes a decision. For him, the real challenge in AI is not just capability. It is making systems auditable, transparent, and boringly trustworthy before anyone bets too much on them.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous part of AI that actually matters: not bigger demos, but better receipts. Everybody loves talking about model capability until a system gets asked to explain itself, or to keep a secret, or to justify a wrong answer without shrugging. Standards and boring infrastructure are not sexy, which is exactly why they’re overdue.
Read more about this at: IEEE Spectrum
