About
I am a second-year Computer Science student, studying at BITS Pilani and Scaler School of Technology. I work on a question that is both technical and philosophical: how should an intelligent system act when it has incomplete grounds for confidence?
At the moment, that question takes the form of research on epistemic self-awareness in language models—using hidden-state signals to predict failure early enough to route a request, retrieve evidence, ask for clarification, or abstain. But the underlying interest is broader: building systems that can represent their own limits instead of concealing them behind fluent output.
The problems I return to
I am more interested in capabilities than in paper-shaped increments. The north star behind my current research is a practical local–cloud routing system: a smaller model should be able to recognise the kind of task in front of it, estimate where its judgment is unreliable, and escalate only when doing so is justified.
That makes the work less tidy than a single benchmark story. It has meant following signals across factual recall, mathematical reasoning, code, calibration, domain transfer, and the places where an apparently promising result stops generalising. I find that more valuable than a clean answer arrived at too early.
A working theory of good systems
I think good systems are legible under pressure. They make their assumptions inspectable, preserve the trail from evidence to decision, and have feedback loops strong enough to correct the model when reality disagrees. That is as true of a research program or a knowledge base as it is of software, organisations, or an individual trying to learn.
This is why I am drawn to systems thinking: not as a taste for abstraction, but as an attempt to see second-order effects, constraints, incentives, and the exceptions that a simple rule misses. I am interested in the gap between a model that works on average and a model that knows when the average no longer applies.
Research as a living record
My work is deliberately cumulative. I maintain a linked research environment where papers, hypotheses, experiments, source material, negative results, revisions, and open questions can remain connected over time. The aim is not to collect notes; it is to make thinking compound.
In practice, the loop is simple: state a claim precisely enough to be wrong, design the test, inspect the result, write what changed, and let the next question grow out of the correction. Writing is part of the method, not a press release after it. If an experiment contradicts an earlier conclusion, I want the correction to remain visible; if a project changes direction, I want the reasons to survive the rewrite. That orientation shapes both the research itself and this site: neither is a polished victory lap, but a record of what I am building, testing, and learning in public.
Elsewhere
I write about technology, agency, learning, and the systems that shape attention. You can find the code behind my work on GitHub, or get in touch by email.