Try the Konami code:
↑↓↑↓←→←→BA
Hi, I'm Naavya, an independent researcher in AI systems and cognition.
Current direction: intrinsic motivation mechanisms to improve reinforcement learning efficiency, exploration quality, and decision reliability.
I'm a recent graduate from the University of Illinois at Urbana-Champaign, where I studied computer science, philosophy, and psychology.
These are my most promising projects. Click to learn more.
Introduces PMS 2.0 (Preprocessing Metacognitive System), a system-agnostic upstream routing layer that decides when to engage, escalate, or refuse tasks before downstream reasoning. Demonstrates improved conditional accuracy, lower compute load, and interpretable governance of computational effort.
Paper exploring metacognitive preprocessing as an upstream system layer — enabling models to defer, gate, or refuse computation to improve conditional accuracy and conserve compute. Inspired by dual-process theories, human metacognition, and a Reddit joke about Star Trek's Replicator.
Exploratory frameworks and prototypes built to think through problems in systems, learning, and representation. These projects are not production tools — they’re attempts to make abstract ideas concrete, test assumptions, and see where models break under pressure.
Current research directions, loosely organized around reliability, efficiency, and decision control in AI.
Studying how humans and machines trade speed for accuracy, how metacognition arbitrates effort when the stakes are real, and how to shift the focus from scaling models to controlling when and how computation is used.
Designing systems that can recognize uncertainty, defer decisions, or reroute reasoning before confidence becomes a liability. Because doubling down is not a strategy - it's a failure mode.
Treating reasoning, reflection, and self-correction as first-class system components, not post-hoc patches. We want systems that surface uncertainty, admit limits, and fail in legible, measurable ways.
Exploring where intuition and heuristics suffice in humans, when explicit reasoning takes over, and how machines can learn when not to reason at all.
For collaboration, feedback on draft work, or shared thinking on metacognition and system design, get in touch.
