Reliable AI agents in software systems
Agent architectures with explicit permissions, validation layers, recovery paths, and execution traces a human can actually read.
My interest is in the reliability of software systems that language models participate in building and operating — not whether a model can produce a working change, but whether that change can be checked, constrained, attributed, and safely rejected.
I came to this from the operational side. Owning a payment gateway teaches you that correctness is not a target but a precondition; running a multi-tenant platform teaches you that isolation failures are found by users rather than by tests. When I began putting language models to work against these same systems, the interesting failures were not wrong answers — they were confident actions taken on incomplete state, and the absence of any record adequate to reconstruct why. Production gave me the problem; it cannot give me the method. That is what I am looking for in graduate study.
AI-assisted software engineering
Evaluating model-generated changes in mature codebases against the criteria engineers actually apply: regression risk, maintainability, security, and operational cost.
Verification and evaluation methods
Detecting incorrect actions, hallucinated assumptions, unsafe tool use, and failure modes that surface only in long-running workflows.
Distributed and cloud systems
System behaviour across APIs, databases, queues, caches, and event-driven services, and how that behaviour changes when an autonomous component sits inside the loop.
Software security and access control
Constraining capability through scoped credentials, authorization boundaries, policy checks, and auditable tool access.