Meet the Anthropic Engineer Keeping Its AI Agent Trustworthy

Gurmeet Saran, an IEEE senior member, works at Anthropic in San Francisco as a member of technical staff, where he is building infrastructure to understand the mechanics behind Claude's behavior and to help ensure the underlying architecture produces reliable outputs. He joined the company in 2025. His stated focus is data accuracy, structure, and trustworthiness, along with keeping private data private.
Saran frames the work around a now-familiar problem in large language model operations: observability. Monitoring databases and web services is a solved discipline with mature tooling, he says, but for LLMs the field is still working out how to see what a model actually did. That gap matters, in his account, because without trustworthy logic behind the outputs, there is no sound basis for claims about model behavior, and he ties that directly to safety implications.
Anthropic's stated mission for Claude, as Saran cites it, is to be genuinely helpful to the people it works with or on behalf of, and to society, while avoiding actions that are unsafe, unethical, or deceptive. Making that verifiable depends on researchers being able to observe and understand how models behave. Saran describes AI agents as especially hard to observe because they take sequences of actions, generating conversational responses with different users over time. Capturing what an agent did, what information it was working from, and whether the outcome matched what the end user expected is the core technical challenge, and it must be done without exposing anyone's private information. Saran says the systems are designed to recognize sensitive data and keep it protected as it flows through, and he calls this a far more demanding observability problem than anything he has worked on before.
He uses a medical analogy to explain the stakes. A doctor monitoring vital signs and making treatment decisions depends on the calibration of the devices producing those numbers; even a slight miscalibration in one device is serious because, in his phrasing, the difficult decisions happen in the margins. His job, by that analogy, is making sure the instruments give accurate readings, so the data means what it says it means.
Saran's path to Anthropic runs through 20-plus years of data infrastructure work. He grew up in Kolkata, spent as much time as he could in the school computer lab, and was struck by the idea that a machine would follow written instructions exactly, without ambiguity. After high school in 1999 he stayed in India, earning a bachelor's degree in electronics and telecommunication engineering in 2004 from the University of Kalyani and a master's degree in software engineering in 2010 from the Birla Institute of Technology and Science in Pilani.
He started his career in 2007 at Yahoo as technical lead of the data team in Bengaluru, while still a graduate student. There he ran into a metric-definition problem that shaped his later work: different teams computed the same website metric, such as daily active site users, with slightly different logic and got conflicting answers. Because leadership used those numbers for resource allocation and product development, not knowing which figure was correct became a real operational problem. The team built a central data repository with defined metrics as a single source of truth, and Saran created the model defining what daily active site users meant, including which events counted and how edge cases were handled. That let teams across the company compute comparable numbers using the same logic. He and his team then built Yahoo! Digits, an internal web analytics tool used across the company's products.
In 2010 he moved to Intuit as a staff engineer in Mountain View, California, helping build analytics infrastructure that unified data across products including TurboTax and QuickBooks. During his time there, Intuit developed a feature to detect fraudulent tax filings in real time, which required building a system that protected taxpayer data while meeting strict U.S. Internal Revenue Service policies. He describes that experience as teaching him that privacy and governance cannot be bolted on after the fact and must be part of the architecture from the start.
Saran joined Meta, then Facebook, in 2016 as a lead software engineer in Menlo Park, California, at a time when the company faced controversy over video metrics; Meta publicly acknowledged an error that overstated views for posted videos. He characterizes that as an architectural problem rather than a coding bug, and built a system defining one governing truth for video view metrics across the site. That system, Metric360, eventually became part of the site's architecture and, by the time he left Meta in 2021, was being used company-wide for video metrics. He then became a principal software engineer at Wealthfront in Palo Alto, California, building the financial services company's data infrastructure.
One throughline across these roles is the single-source-of-truth pattern: define a metric once, encode its edge cases, and force every downstream consumer to use that definition. For AI agents, the analogous problem is harder, because the "metric" is a sequence of actions and reasoning steps rather than a countable event. This suggests why Saran frames his Anthropic work as instrumentation rather than reporting: capturing traces of agent behavior, the context a model worked from, and outcome alignment, while applying sensitive-data protections inline.
The privacy constraint is the part that distinguishes this from conventional observability. Standard tracing stacks log prompts and tool calls liberally; agent traces can contain user data, so collection, storage, and access controls have to be designed together rather than retrofitted. A likely trade-off is granularity: richer traces make behavior easier to audit but enlarge the sensitive-data surface, pushing teams toward redaction, aggregation, or access-scoped retention. Saran's claim that privacy must be architectural maps onto that tension.
Saran also joined IEEE in March 2025, motivated by a desire to contribute through peer review, publishing, and engagement with the broader engineering research community. He argues IEEE has a role in shaping AI development through standards work, drawing a parallel to building codes and software testing standards, and points to access as a second area: IEEE operates in many countries where engineers lack easy pathways into the global professional community, which he says matters as AI development spreads beyond a handful of organizations in a few cities.
Why it matters: teams shipping agents increasingly need audit trails that explain what a system did and why, not just dashboards of latency and cost, and this account highlights how hard that is when traces contain user data. Developers building agent platforms should expect the observability and privacy concerns to be coupled, and treat trace design as an architectural decision rather than a logging afterthought. Saran's view that auditability is under-discussed relative to model capabilities is his own assessment, and the extent to which Anthropic's internal tooling will generalize beyond Claude is not stated.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
If you’ve ever wondered how Claude and similar AI agents can answer complex questions in seconds and where they get the information, you’re not alone. Gurmeet Saran works to find that answer every day as an engineer at Anthropic in San Francisco, the AI research company behind Claude. The IEEE senior member is developing an infrastructure to understand the mechanics behind the chatbot’s behavior and to ensure its underlying architecture produces reliable outputs. Gurmeet Saran Employer Anthropic in San Francisco Ti