Sasank Aramadaka.
From testing bank software to governing AI agents, the engineer's two-decade career and seven-year research record trace how financial services learned to trust automation.
Banks are handing more of their customer conversations to artificial intelligence. Voice bots answer calls, agents look up accounts and generative systems draft replies in seconds. The question that now keeps compliance teams awake is simple: when an AI agent makes a decision, can anyone prove why?
Sasank Aramadaka has made that question the focus of his latest research. His newest study, published in August, sets out how AI agent decisions in financial services can be governed and traced from start to finish. It is the sixteenth research article he has published since 2019, and it caps a body of work that runs from the plumbing of bank data systems to the behavior of autonomous agents.
Aramadaka, a lead AI/ML engineer with more than 20 years in technology, builds generative AI voice agents for the banking contact center. His research and his day job have grown up together.
THE LATEST: WATCHING THE WATCHERS
Aramadaka's three most recent studies form a clear sequence. Last October he examined multi-agent generative AI workflows for contact centers, cataloguing design patterns that work and the failure modes that appear when several agents hand tasks to one another. In March he turned to evaluation itself, studying how prompts are designed and how enterprise LLM agents are assessed, including the risk that the methods used to judge AI are themselves flawed.
The August study completes the picture. It argues that autonomy in banking must be auditable by design: every decision an agent takes should leave a trail that shows what it was asked, what information it used, which tools it called and why it acted as it did. Without that trail, he contends, regulated institutions cannot responsibly let agents act on customers' behalf.
FROM TESTING SOFTWARE TO TESTING INTELLIGENCE
The roots of that thinking go back years. Aramadaka spent much of his career in quality engineering for banking, mortgage finance and healthcare systems, leading teams that automated UI and API testing with Selenium, Java, TestNG and JMeter, and validated ETL pipelines built on Informatica and Ab Initio.
His first studies reflected that work, examining how to hold gateway services to API contracts and how to validate microservice systems end to end in regulated settings. Later research tackled data integrity at scale: comparing data across different database engines, verifying cloud warehouses layer by layer, linking batch schedulers to automated comparisons and keeping financial reference data accurate.
As a senior cloud architect, he worked across Amazon S3, Redshift, EMR, Lambda, Glue and Step Functions, and his research followed him into cloud migration, covering cutover and rollback planning and the security of capital markets data during warehouse moves.
INTO THE AGE OF VOICE AND AGENTS
In 2024 his research shifted to conversational AI. He studied how speech markup, pauses and intonation shape the customer experience of voice bots, how backend service layers for voice agents should be designed and documented, and how banks can govern interaction design so every conversational path is approved by the business before it goes live.
In 2025 he went deeper into the engineering. One study showed how selective suppression of entries in a vector index can reduce confusion between similar customer intents. Another set out how to design guardrails for autonomous agents in regulated banking.
Each study maps onto his current work. He designs conversational flows in Dialogflow CX, deployed a vector embedding and retrieval pipeline on Google Cloud Run with Vertex AI Vector Search, builds REST services that connect voice agents to core systems, uses SSML to make bots sound natural, and tracks agent performance through Splunk, BigQuery and Looker Studio.
WHY IT MATTERS
Aramadaka's contribution is a rare continuity. The same discipline he applied to proving that a warehouse table matched its source, he now applies to proving that an AI agent behaved correctly. That continuity matters because the industry is learning, sometimes the hard way, that AI needs the same rigor as any other critical system, and more.
The impact is felt most in regulated finance, where an unexplained decision is a liability. His work gives banks practical methods to improve accuracy, constrain agent behavior, evaluate results honestly and keep a complete record of what their AI did. Those lessons apply equally to insurers, healthcare providers and any organization putting AI in front of customers.
As AI agents take on more responsibility, the people who know how to verify them will matter as much as those who build them. With 16 studies and two decades of hands-on experience, Sasank Aramadaka has positioned himself firmly among them.
