About
I’m a Researcher at Qualcomm AI Research, working on generative modeling, reinforcement learning, and representation learning, including post-training for reasoning models (on-policy distillation, RL fine-tuning) and, increasingly, embodied world models.
One line of work I keep returning to is DDIL, where I reframed diffusion distillation as imitation learning: the diffusion student is a policy, its denoising trajectory is a rollout, and teacher/student mismatch is a covariate-shift(exposure-bias) problem. The correction is DAgger-style mixed-distribution training, where the model learns on its own induced states while preserving teacher and reference states that maintain diversity. This work was instrumental in the world’s first sub-0.6s on-device Stable Diffusion deployment. The transferable idea is not “faster sampling.” It is amortizing a slow teacher or search process into a faster model without collapsing the diversity of the pretrained prior, a structure that reappears in reasoning post-training, verifier-guided learning, and embodied world models.
A thread I find especially interesting is competence as a runtime signal: a verifier or confidence head cannot recover competence-relevant structure the representation has already discarded. So the leverage tends to be upstream, at the representation, rather than at a better score on top of a fixed one.
15+ peer-reviewed publications at JMLR, NeurIPS, ICML, CVPR, and ECCV →
Highlights
World’s first sub-0.6s Stable Diffusion on mobile. ML lead in a cross-functional team at Qualcomm delivering the world’s first and fastest on-device text-to-image generation. Built the training pipeline from scratch, managed 50TB data pipelines via MosaicML MDS, identified the manifold-thresholding technique and modified diffusion sampling for on-target quality, and led debugging across the deployment stack. Covered by The Verge and Engadget, featured at MWC’23 and Snapdragon Summit.
Diffusion distillation as imitation learning (DDIL). Identified trajectory diversity collapse as a fundamental failure mode in distilled diffusion models and proposed a DAgger-style on-policy correction that trains on a mixture of student-induced states (correcting covariate shift) and teacher/reference data (preserving diversity), recovering intermediate marginal distributions. First applied to progressive distillation (behavior cloning) to enable the sub-0.6s on-device deployment, then extended to distribution matching for larger models.
Reliable representations under distribution shift. Graduate research at Oregon State (advisors: Tom Dietterich, Alan Fern) on open-set recognition and failure detectability. I studied what a representation preserves about the data manifold, where competence-relevant information is lost, and why a downstream classifier, confidence head, or verifier cannot recover structure the representation has already discarded. Pairing discriminative objectives with generative priors raised the ceiling on detectable failures.
Hardware-first grounding. Before ML: satellite fault-tree analysis (ISRO collaboration), then founding engineer on Nino, a consumer humanoid robot, covering DoF allocation, static load analysis, EKF state estimation, and 3D-LIP gait control, from concept to walking prototype.
Current Research Direction
A loop I’m interested in studying across settings:
- Search creates supervision. A model improves by exploring its own generated computation: rollouts, verifier-filtered samples, self-consistency, process supervision, and latent futures, not only by imitating fixed targets.
- Distribution-aware correction / on-policy distillation internalizes it without collapse. Once a model learns from its own behavior, the training distribution becomes endogenous; the model must learn on its induced states while preserving enough reference support to keep diversity, coverage, and recoverability.
Much of my current attention is on reasoning: search amortization, on-policy distillation, verifier-guided learning, and competence-gated test-time compute, because it gives fast, measurable feedback through verifiers, critics, search traces, and on-policy rollouts, with clear failure modes such as critic and verifier overfitting and coverage collapse. I’m interested not only in whether a recipe improves reward, but in why and how: whether it preserves coverage and shifts measurable quantities such as rollout diversity, teacher/student overlap, feature alignment, and credit localization.
I’m interested in how the same loop carries to embodied world models, where it appears over future states via an internal world model and action rollouts under physical constraints. One key aspect here is the representation, interface, and competence-monitoring layer between high-level reasoning and low-level control. DDIL and the reasoning work are the concrete starting points; the multimodal and embodied directions are where I’d like to see the same ideas carry.