← Back to Home

Projects

Systems. Robotics. Hardware. AI

TinyEdge: ML Compiler →

An end-to-end ML compiler that lowers MobileNetV2 through 4 IR layers with custom CUDA kernels, delivering a 4.67× inference speedup over PyTorch eager.

CUDA C++ PyTorch TVM MLIR LLVM

• Built an end-to-end ML compiler that lowers trained PyTorch MobileNetV2 through 4 IR layers (torch.fx → TVM Relax → TIR → LLVM/CUDA) plus a custom MLIR dialect, delivering a 4.67× inference speedup (29.4 → 6.3 ms) over PyTorch eager.
• Hand-wrote 4 CUDA kernels (matmul, tiled matmul, parallel reduction, vector add) and shipped them into PyTorch as fused custom ops (Linear+ReLU, GlobalAvgPool) via C++/PyBind11, achieving 6.9× matmul on T4 (39.7 → 5.78 ms); TIR schedules hit 63× on a 1024³ matmul (1320 → 21 ms, 104 GFLOP/s).
• Added a graph-transformation layer: Conv+BatchNorm folding, INT8 post-training quantization with per-channel calibration, and an operator partitioner that auto-assigns all 141 ops.

---

Advanced Out-of-Order RISC-V Processor →

Developed a high-performance RISC-V processor featuring advanced pipelining, branch prediction, and caching mechanisms.

RISC-V SystemVerilog Processor Design Computer Architecture

• Developed a high-performance, pipelined processor in SystemVerilog, implementing the RISC-V ISA with features like R10K-style register renaming, N-way superscalar execution, early tag broadcast, and GShare-based branch prediction.
• Designed advanced memory subsystems, including mechanisms for instruction prefetching, non-blocking instruction and data caches, as well as a victim cache for improved cache performance.
• Engineered a load-store queue with forwarding to optimize memory operations and minimize pipeline stalls.
• Validated using a suite of benchmark programs, including Neural Networks, sorting algorithms etc with an average CPI of 1.8

---

Enfield Engine →

A high-performance, multi-threaded game engine crafted in C++ with Lua scripting and advanced physics.

C++ Game Engine Box2D SDL Lua ImGui

• A high-performance, multi-threaded game engine crafted in C++, integrating Lua scripting, and a PubSub System.
• Implemented advanced 2D physics features by incorporating the Box2D middleware, leveraging SDL for rendering, and a Graphical User Interface using ImGui.
• Enhanced engine efficiency and performance through Data-Oriented Design and Spatial Hashing techniques.
• Feel free to check out the source code on GitHub.

---

Roll Over Beethoven: A Cloth Piano →

A portable cloth piano using capacitive cloth and pressure-sensitive gloves.

C Embedded Systems Audio Synthesis Real-time Processing

• Designed a portable piano using capacitive cloth, pressure-sensitive gloves and a sustain pedal.
• Developed a C tutorial for beginners.
• Wrote drivers for the sensors and display. Synthesized different sound timbres for diverse playing modes.
• Feel free to check out the source code on GitHub.

---

Robotics →

Algorithms for motion planning, kinematics, and control systems in robotics.

Robotics Path Planning Kinematics Control Systems PyTorch

• Motion Planning & Kinematics:
  • Implemented an A-star path planning algorithm to find collision-free paths in 2D environments.
  • Developed Euler and Velocity Verlet integrators, a PID controller, and a simple pendulum simulation.
  • Built forward kinematics for arbitrary robots, predicting the spatial layout in a 3D world based on joint configurations.
  • Designed inverse kinematics for end-effector position control using gradient descent with the Jacobian transpose.
  • Created collision-free motion planning algorithms using RRT, RRT*, and RRT-Connect to navigate complex environments.
• Control Strategies: Implemented an iterative linear quadratic regulator (iLQR) for inverted double-pendulum swing-up on a cart; developed shooting and collocation methods for finite horizon Linear Quadratic Regulation (LQR) with input constraints; applied MPPI for cartpole swing-up, optimizing nonlinear trajectories and achieving stable pole-up configurations.
• Reinforcement Learning: Designed reward functions for robotic manipulation, optimizing task success rates in obstacle-free and obstacle-present environments.

---

Operating Systems →

Custom system calls, Unix shell, and memory management in C++.

Operating Systems C++ Linux Kernel

• Custom system call in the Linux kernel, utilizing C and VirtualBox.
• Custom Unix Shell, a focus on core functionalities like command parsing and execution.
• Multi-CPU supported thread library.
• Kernel-level pager to manage and translate virtual address spaces using system calls.
• Multi-threaded server for hierarchical file systems managing client interactions via network messages with robust security against untrusted inputs.
• Developed an assembler, simulator, multiplier, linker, pipelined datapath simulator, and data/memory cache simulator.

---

Deep Learning →

Implementation of neural networks and reinforcement learning algorithms in Python.

Deep Learning Neural Networks Computer Vision PyTorch Transformers

• Neural Networks: Implemented fully connected and convolutional neural networks; developed and trained one-stage (FCOS) and two-stage (Faster R-CNN, integrating a Region Proposal Network) object detectors; created image captioning models using RNN, LSTM, and Transformer networks with spatial attention; implemented generative adversarial networks, network visualization techniques to create saliency maps, and style transfer.
• Control Strategies: Implemented an iterative linear quadratic regulator (iLQR) for inverted double-pendulum swing-up on a cart; developed shooting and collocation methods for finite horizon Linear Quadratic Regulation (LQR) with input constraints; applied MPPI for cartpole swing-up, optimizing nonlinear trajectories and achieving stable pole-up configurations.
• Reinforcement Learning: Designed reward functions for robotic manipulation, optimizing task success rates in obstacle-free and obstacle-present environments; created a Variational Autoencoder (VAE) and applied Gaussian Process regression using GPyTorch with various kernels to model noisy data with uncertainty.
• Robotics: Implemented an A-star path planning algorithm for collision-free paths in 2D environments; developed Euler and Velocity Verlet integrators, a PID controller, and a simple pendulum simulation; built forward and inverse kinematics algorithms for 3D robot configurations; designed collision-free motion planning using RRT, RRT*, and RRT-Connect algorithms.

---

Networking →

Video CDN implementation and network diagnostic tools.

Networking Distributed Systems C++ C

• Implemented a Video CDN with adaptive bitrate streaming and DNS load balancing.
• Developed WTP protocol over UDP for reliable data transfer handling packet loss, delay, and corruption.
• Built a static-routing router to enable real-world packet routing, network diagnostics with operations like ping, traceroute, and HTTP file download.

---