Projects

RL-Based Quadruped and Humanoid Control

Developed PPO-based reinforcement learning pipelines using MJX and Brax for learning locomotion in quadrupeds and humanoids. Built robust checkpointing, parallelized training, and evaluation tools to analyze stability and performance.
Quadruped Demo Humanoid Demo
  • Built PPO pipelines in JAX + Brax with checkpointing and evaluation
  • Integrated with MJX physics simulation
  • Generated video/HTML visualizations for policy testing
View Project →

CUDA GPU Programming Portfolio

A suite of CUDA-accelerated programs built for high-performance computing experiments. Implemented low-level GPU kernels and benchmarked performance against SIMD-optimized CPU versions (AVX2/SSE).
One of my first kernels
  • Custom CUDA kernels for matrix ops, softmax, image filters
  • Memory coalescing, shared memory tuning, and kernel optimization
  • Built custom CUDA conv2D kernels on image data sourced from OpenCV
View Project →

Visual Navigation & 3D Mapping

Developed a monocular visual navigation game with SLAM-like behavior. Integrated camera and IMU data for real-world TurtleBot3 mapping and tested with multiple sensors including 360° camera and LiDAR.
Navigation Demo
  • Feature-based navigation using SIFT and OpenCV
  • SLAM-based mapping with IMU + LiDAR integration
  • Adapted for TurtleBot3 with ROS-based deployment
View Project →

Hardware-Software Co-Design for RISC-V

Built a 5-stage pipelined, cycle-accurate RISC-V simulator in Python. Designed custom RISC-V instructions for Transformer acceleration, and achieved 8× speedup through hardware-software optimization.
RISC-V Simulator
  • Implemented hazard resolution using forwarding/stalling
  • Added MAC, softmax, and saturating vector ops
  • Accelerated Transformer inference in simulation
View Project →

Stock Price Forecasting with Sentiment Analysis

Developed a forecasting model combining news sentiment and technical indicators. Used DL models like Transformer, DP-LSTM, and CNN-LSTM, outperforming traditional statistical methods on real-world data.
Stock Forecasting
  • Processed 6,000 news articles with sentiment analysis
  • Integrated 15+ technical indicators with price data
  • Ran models on NYU’s HPC (Slurm + Singularity)
View Project →