About
Ph.D. Student, Computer Science and Software Engineering at Auburn University
AI/ML Researcher & Engineer
Biography
I am a Ph.D. student in Computer Science and Software Engineering at Auburn University, where my research sits at the intersection of empirical evaluation and applied machine-learning systems.
Two questions run through most of my work. The first is a measurement question: when we say one model or one representation is better, what exactly have we measured, and would that conclusion survive a different dataset, a perturbed input, or a significance test? That question produced a systematic zero-shot evaluation of 41 open-weight language models across eight intent classification datasets, where the interesting result was not a leaderboard but the finding that several widely used benchmarks no longer discriminate between current models at all.
The second is a systems question: what does a model actually need to remember, retrieve, or represent in order to be useful under real constraints? That question runs through a controlled study of interaction representations for early violence detection in video, a retrieval-augmented system over roughly 25,000 research papers, and an incident-response agent whose memory lives in a database rather than a context window.
Before returning to research I spent about eighteen months as an Associate Software Developer at Emids Technologies in Bangalore, building CRM modules and REST services in Java and Spring Boot. That period is why I care about systems that are maintainable and honest about their own failure modes, not just accurate on a held-out split.
How I approach research
My working assumption is that the measurement is the contribution. A new architecture that improves a number is only interesting once you can say which part of it did the work, and whether the improvement would survive a different dataset, a perturbed input or a significance test. Both of my first-author papers are structured around that question rather than around a leaderboard position.
That leads to three habits. I prefer controlled comparisons to broad claims — hold everything fixed and vary one thing, then report the result scoped to the conditions it was measured under. I try to make systems honest about their own limits: labelling which implementation is actually running, keeping a raw similarity score rather than rescaling it to look impressive, and writing the limitations section before the abstract. And I treat negative and deflationary results as findings — a benchmark that has saturated, or two models that turn out to be statistically indistinguishable, is worth reporting precisely because it changes what other people should do next.
The applied side of my work exists for the same reason. Building a retrieval pipeline over 25,000 papers, or an agent whose memory is a transactional database, surfaces constraints that do not appear in an offline evaluation — and those constraints are usually where the interesting research questions are.
Current goals
- Take the evaluation protocol from the 41-model study beyond single-label intent classification, into tasks where label ambiguity is intrinsic rather than accidental.
- Report the contribution of interaction representations as a function of how much of an event has been observed, rather than as a single number at a single observation ratio.
- Move the AUSME retrieval system from a working pipeline to something faculty and research groups actually query.
- Keep building systems where the memory, the retrieval and the human decision are part of one auditable path.
Collaboration
I am glad to hear from people working on evaluation methodology, video understanding, retrieval-grounded systems, or applications where a model's reliability matters more than its peak score. I am currently open to:
- Research collaborations
- Research internships
- Ph.D.-level co-authorship
- Reviewing and programme-committee service