I am an AI researcher with a background in systems engineering.
I care about how models work from the ground up. Most of my work involves building tensor engines and autograd from scratch in C++ and CUDA, training diffusion models for video frame interpolation, and experimenting with reinforcement learning on planning problems.
- Tensor Engines & Compute: Building an autograd engine and tensor library from scratch in C++ and CUDA (
MyTorch), focusing on computation graphs, custom kernels, and GPU memory layouts. - Diffusion & Frame Generation: Working with diffusion models and flow matching for video frame interpolation, generating in-between frames from past and future context.
- Reinforcement Learning & Planning: Testing RL algorithms on structured environments like BlocksWorld with sparse rewards, studying state representation and credit assignment.
|
A deep learning tensor library and autodiff engine built from scratch in C++ and CUDA to understand GPU execution and memory management directly.
|
Using diffusion models to generate intermediate video frames from surrounding frames, handling non-linear motion and occlusions.
|
|
Training RL agents to solve structured planning tasks like BlocksWorld where rewards are sparse and actions depend on object relationships.
|
Clean, minimal implementations of foundational papers to understand the details that matter during training.
|
Current Literature Queue & Completed Deep Dives (Click to expand)
| Topic | Paper / Resource | Focus & Notes |
|---|---|---|
| Diffusion Models | Denoising Diffusion Probabilistic Models (Ho et al.) | Derivation of the variational bound and noise schedules |
| Flow Matching | Flow Matching for Generative Modeling (Lipman et al.) | Straight vector fields and optimal transport paths |
| Video Generation | Diffusion Models for Video Generation & Interpolation | Temporal cross-attention and conditioning on surrounding frames |
| RL & Planning | Reinforcement Learning in Relational Domains | State representations and credit assignment in BlocksWorld |
| GPU Architecture | Programming Massively Parallel Processors (Kirk & Hwu) | Memory coalescing, shared memory, and warp divergence |
"Curious about how machines learn structure, not just patterns."