MS CS (Machine Learning) @ Columbia University
I build systems that make ML inference faster. My research interests are in efficient LLMs training and inferece, computer vision and robotics.
batched_specdec — Batched Speculative Decoding Engine
PyTorch implementation with prompt batching, non-uniform acceptance lengths, and parallel draft verification.
distillSpec — On-Policy Distillation for Speculative Decoding
Achieved 5% token acceptance gain on GSM8k via Reverse KL distillation. Includes batched verification, KV-caching, and pruning.
→ HuggingFace Models
3DEgoACT — Viewpoint-Invariant Robot Manipulation
Fuses PointNet 3D encoding with egocentric vision. 70% zero-shot success on perturbed camera views where baseline ACT fails.
→ Dataset
Stack: Python, C++, PyTorch, HuggingFace, CUDA, MuJoCo