Research

When does misalignment emerge during RLVR?

BlueDot Impact, Technical AI Safety Project · 2026 – ongoing

Reinforcement learning with verifiable rewards (RLVR) is the training recipe behind today's reasoning models, but nobody really knows when (or how) misaligned behavior appears along the way. My approach is simple: take intermediate checkpoints of an open-weight model (OLMo-3), evaluate each one, and map where the misalignment starts relative to the capability gains. I am also training linear probes on the checkpoint activations, to test whether the internal signals of misalignment appear before the behavior does. If they do, that gives a cheap early-warning monitor, i.e. a way to flag problems before they show up in the outputs.

Optical Ising machine for physically solving PDE dynamics

Sandia National Laboratories, Center for Integrated Nanotechnologies · 2026 – ongoing

Simulating the dynamics of nonlinear partial differential equations gets expensive quickly on digital hardware. I am building an optical optimizer (an Ising machine) that physically solves the dynamics of the Cahn–Hilliard equation, the workhorse PDE of phase separation. Instead of numerically stepping through the equation, the optics itself does the computation. For large systems, this can potentially give a ≥9× gain in both speed and energy cost.

Near real-time faint-object detection on GPUs

Zetascope LLC · 2026

Detecting faint moving objects is hard when the per-frame signal sits below the noise floor; you have to integrate the intensity along millions of candidate trajectories, which is a brute-force compute problem. I developed a low-power, field-deployable detection pipeline that runs in near real-time on quantum shot-noise-limited images. I hand-optimized the CUDA C++ kernels for the NVIDIA Blackwell GPU, using memory-hierarchy and roofline analysis to guide the optimization, and cut the runtime by ~108×.

Deep-microscopic optical band theory of chiral tellurium

Purdue University, Jacob Group · 2024 – 2025 · Phys. Rev. B Letter, Editor's Suggestion

Tellurium's optical activity is known experimentally to diverge near the band edge, but decades of first-principles attempts failed to reproduce it. The reason is that optical gyrotropy is intrinsically a finite-momentum effect. I built the first fully quantum, lattice-resolved optical band structure of chiral tellurium, solved the dispersion problem with quantitative agreement against experiment across the mid-infrared, and found hidden optical waves and optical spin textures inside the lattice.

Alongside the theory, I co-authored Purdue-PicoMax, an open-source C++ package for nonlocal optical response in crystals. It reaches ~5× lower error than existing methods, and its predictions were later validated by electron-energy-loss measurements.

Quantum theory of structured light

Purdue University, Jacob Group · 2022 – 2024 · New J. Phys. & JOSA B

I settled a long-standing debate on the correct quantum commutation relations for the angular momentum of light; at the center of it sits a widely propagated textbook error. The tutorial I wrote on this was JOSA B's most-downloaded article for three consecutive months. I also built the first quantum field theory of spatiotemporal optical vortices beyond the paraxial limit, which predicts a previously unrecognized, experimentally accessible quantum fluctuation in their orbital angular momentum.

Quantum sensing with NV centers in diamond

Purdue University, Jacob Group · 2021 – 2022 · multiple co-authored articles

I built experimental infrastructure for single spin-qubit control and readout: I co-designed a cryogenic confocal microscope for optically-detected magnetic resonance, created shallow NV defects (~30 nm from the surface) via ion implantation and annealing, and developed a GHz phase/amplitude modulation setup for spin-textured light using temperature-controlled fiber Sagnac interferometry. This infrastructure underpins published measurements of photonic spin texture with NV-center quantum sensors.

Microwave resonators for cryogenic spin spectroscopy

IISER Kolkata, Mitra Group · 2019 – 2021 · J. Phys. D & IOP SciNotes

I designed (CST Studio), fabricated (UV photolithography), and characterized two on-chip microwave resonators: a short-circuited coplanar waveguide, and a localized spoof-surface-plasmon structure. These were the field-concentrating elements of home-built electron spin resonance spectrometers, operating from room temperature down to 10 K.