Active learning (data selection) method (CMAP) that uses neural collapse geometry to pick what to label: it scores each unlabeled sample by how much it would perturb the learned class geometry. Reaches strong accuracy with about 20% fewer labels on ImageNet, and is more robust to label noise and more generalizable to OOD data and novel classes. Validated on OOD detection, long-tailed data, transfer learning, and novel-class discovery.

Training-free, annotation-free pipeline that composes 2D vision-language foundation models with classical geometry (camera models and LiDAR) for open-vocabulary 3D detection. Also introduces a robustness benchmark with physics-based fog simulation and unreliable LiDAR, which exposes severe failures under sensor and distribution shift.

Neurosymbolic system for the ARC-AGI benchmark that reaches 41% accuracy, built around LLM test-time adaptation. Symbolic program synthesis augments each task's examples, and a meta-learned base LLM is fine-tuned with QLoRA on those augmented instances (self-supervised). Inference uses a custom DFS over candidate token sequences with pruning heuristics that enforce spatial-reasoning constraints, at the same time complexity as greedy sampling.

Can deep RL agents acquire core knowledge priors and compose them across tasks? I trained PPO agents (custom CNN policy) on a curriculum of manipulation tasks in procedurally generated, ARC-inspired grid worlds, bootstrapped with imitation from a BFS planner. Finding: agents learned task-specific mappings but did not transfer latent task structure across environments. Joint online RL and offline distillation sped up convergence within a task, but did not fix cross-task generalization. Also explored meta-learning, test-time adaptation, inverse RL, pretraining, and open-ended learning. Advised by Prof. Gautam Shroff.

Semi-supervised active learning framework for object detection that combines foundation models with human-in-the-loop annotation. Uses Neural Collapse to build a targeted acquisition function, which enables efficient training under tight annotation budgets and strong performance with minimal human supervision.

Implemented a distributed key-value store on top of a from-scratch implementation of the Raft consensus algorithm. Includes a custom leader lease implementation that ensures fault tolerance and log consistency across distributed nodes.

Wrote a rendering engine from scratch in raw OpenGL for real-time global illumination. Implemented a custom dynamic voxelization algorithm with voxel cone tracing for indirect lighting, achieving real-time performance while maintaining visual fidelity comparable to offline rendering techniques.