← All speakers

Anirudh Veeraragavan

Conference affiliation: Fieldguide

Anirudh Veeraragavan is the lead security engineer and first security hire at Fieldguide, which builds software for audit and advisory firms. His work spans shared cryptographic infrastructure at Plaid and controls for financial agents at Fieldguide. In both settings, he explains the engineering beyond a correct result: getting applications onto a shared security service, preserving a reviewer’s signed decision, and recording enough evidence to reconstruct an agent’s work.

On this page

Building and adopting shared security infrastructure

Anirudh studied at UCLA from 2016 to 2020. An early project, NXTPitch, combined machine-learning services to give students feedback on pitches and application materials. He developed it with Rahul Sheth. For its 2017 machine-learning extension, Anirudh wrote the Python backend, integrated services from Microsoft Azure, IBM Watson and Eyeris, and built the logic for extracting content from videos and documents.

He subsequently spent five years at Plaid, most recently as a Staff Software Security Engineer. He led product-security initiatives, including a workstream developing agentic penetration-testing systems.

Anirudh also worked on Plaid’s key-management system. In a technical account co-authored with Shuaiwei Cui, he described the team’s decision to centralize cryptography that services had implemented differently. The system retained AWS KMS as its root of trust while moving routine operations into infrastructure tailored to Plaid’s scale and access controls. Envelope encryption let applications encrypt payloads locally and send only their data keys through the central service, reducing traffic and making performance more predictable.

Completing adoption was a substantial part of that work. The team tackled challenging services first, provided tools and documentation for most migrations, then worked directly with owners of the remaining exceptions. Their account explains why building the central service was only part of the task: its security benefits depended on moving applications onto it and continuing to operate it.

Making financial-agent decisions inspectable

At Fieldguide, Anirudh argues that an experienced auditor new to an engagement should be able to follow an agent’s work: understand the test, examine the evidence and reconstruct the conclusion. His writing on defensible agents and financial-agent talk explain why a strong aggregate evaluation score cannot establish whether a particular audit decision was justified.

His approach connects several engineering decisions:

  • Preserve human-reviewed work. In his bank-reconciliation example, a reviewer flags an outstanding check and signs the finding. New evidence may correctly resolve the discrepancy, but an agent that overwrites the finding changes what the signature appears to approve. Anirudh argues that authorization must account for the record’s state: the agent can draft unreviewed work as itself, while a correction to signed work must remain a suggestion for a reviewer to consider.
  • Record execution at the level of individual steps. Fetching a document, extracting a field, matching records and reaching a conclusion should leave identifiable steps with recorded inputs and outputs. Anirudh distinguishes deterministic operations performed by code from model-based extractions and judgments. When two runs disagree, an investigator can compare and replay the step where they diverged instead of rerunning the entire process and receiving another unexplained answer.
  • Identify the evidence precisely. A citation to “January bank statement, page two” can point to the wrong account or leave the document version unclear. Anirudh calls for a document identifier, version and content hash that establish which contents the agent used. That identity also makes evidence rules enforceable—for example, requiring a statement obtained directly from the bank rather than one uploaded by the client.
  • Audit the evaluation itself. Previously reviewed audits can contain mistakes, depend on evidence outside the recorded work, or reflect superseded rules. In Anirudh’s illustrative evaluation, four of nine apparent misses came from incorrect expected answers. Twelve other decisions matched the expected answers but were wrong under revised business rules or audit methodology. A reported score of 97 percent concealed both problems. His example gives reviewers a reason to inspect apparent successes alongside failures and ask which evidence and rules justify each label.

These controls also determine when the agent should stop. Anirudh identifies attempts to change reviewed work, disagreements between runs and evidence that fails the required checks as reasons for human investigation. He favors selective escalation with a clear discrepancy and enough context to examine it. Making a case easy to verify can still leave substantial work before approval; repeatedly asking reviewers to click yes risks turning their signatures into automatic responses.

1 conference talk

Loading more talks and highlights…

References