Anirudh Veeraragavan

About Anirudh Veeraragavan

Anirudh Veeraragavan is lead security engineer at Fieldguide, a vertical AI company that builds agents for audit and advisory firms. His work covers the security and governance of agents that reach conclusions in financial audits, where the firm's output is the conclusion itself and must survive an external inspection.

In an AI Engineer talk, he uses a bank reconciliation agent to show why good metrics and happy users are not a defense when an inspector asks how a specific decision was reached. He borrows the audit standard that an experienced auditor new to the engagement must be able to trace the test, the evidence and the decision, and translates it into four engineering controls.

Notable Positions

On Authorization

Permission is per object and per state. An agent may draft unreviewed lines in its own name, but it cannot overwrite work a human has signed, because the signature would then sit on content the reviewer never saw. See Human-in-the-Loop.

On Recorded Runs

Logs of inputs and outputs do not explain why two runs of the same procedure disagree a year apart. Each step needs an ID, inputs, outputs and a marker of whether it was deterministic code or a model call, so individual steps can be rerun and compared.

On Evidence Provenance

A citation like "January bank statement, page 2" is a string, not evidence. It should carry a document ID, version, content hash and origin, which lets a firm enforce admissibility rules such as accepting only statements pulled directly from the bank.

On Eval Integrity

A benchmark built from verified prior-year audits can still be wrong, because items get superseded and methodologies change. In his example, a 97% score hid four mislabeled misses and twelve stale passes. See Agent Evaluation.

Key Quotes

• "In an audit, the conclusion is the product." (on why process matters more than being right)
• "A link is an identifier, not just a line of text." (on evidence provenance)
• "These escalations should be easy to verify, but not necessarily easy to approve." (on avoiding automation bias)

Related Reading

• Human-in-the-Loop - Escalation design that resists rubber-stamping
• Agent Evaluation - Why the eval set itself needs auditing