Overview

This repository showcases systems-focused and performance-oriented work across C++ engine architecture, GPU programming, and applied deep learning.

The projects emphasize:

• Low-level system design
• Parallel computation
• Scalable data pipelines
• Real-time performance and efficient resource management

---

Projects

GPU Animation Engine

C++ · DirectX 11 · HLSL

A real-time 3D animation engine that offloads skeletal animation and hierarchy computation entirely to the GPU using compute shaders.

Key Highlights:

• GPU-based skeletal animation (no CPU bottleneck)
• Multi-stage compute shader pipeline (blend → hierarchy → world matrices)
• Explicit GPU resource management (SRV / UAV / CBV)
• Engine architecture using manager + state patterns

---

Optimized C++ -- Performance Engineering

C++ · SSE Intrinsics · MSVC

A collection of performance engineering assignments demonstrating low-level memory management, cache optimization, SIMD vectorization, custom heap allocators, and expression templates -- with measured speedups across all implementations.

Key Highlights:

• Custom heap allocator with next-fit allocation, block subdivision, and bidirectional coalescing via secret pointers and bit-packed headers
• Cache-optimized hot/cold data separation achieving 40x speedup over naive linked list traversal
• SIMD intrinsics (SSE) for 4x4 matrix multiplication, vector-matrix products, and LERP
• Expression template proxy objects eliminating intermediate temporaries for 2.3x speedup
• Hybrid merge/insertion sort on linked lists (381x faster than pure insertion sort)
• Implicit conversion prevention via private template constructor poisoning

---

Concurrent Maze Solver

C++17 · std::thread · std::atomic · MSVC

A multithreaded maze solver that runs two depth-first searches concurrently from opposite ends of a shared std::atomic_uint[] cell array. Achieves ~2.7× speedup over an optimized single-threaded DFS (and ~3.8× over BFS) on mazes up to 20,000 × 20,000 cells (400 million cells), with synchronization reduced to a single atomic instruction per cell visit.

Key Highlights:

• Lock-free meet-in-the-middle detection via fetch_or -- one atomic RMW per cell visit, no mutexes or fences in the hot path
• memory_order_relaxed synchronization throughout, leveraging algorithmic idempotence under stale reads
• Visit bits co-located with wall data in a single std::atomic_uint -- one cache-line touch per step
• Corridor-skipping DFS that pushes only branch points onto the choice stack, not every cell
• Pre-reserved choice stacks (reserve(400000)) and result vectors -- zero allocator pressure inside the timed region
• Wall-clock benchmarking with seq_cst fence brackets around QueryPerformanceCounter to prevent reordering bias

---

Image Captioning System

Python · TensorFlow

A multimodal deep learning pipeline combining CNN encoders and attention-based sequence decoders to generate image captions.

Key Highlights:

• CNN + RNN (LSTM) with attention mechanism
• End-to-end training and evaluation pipeline
• Techniques for improving generalization and stability
• Handling vanishing gradients and overfitting

---

Real-Time 3D Rendering Engine

C++ · DirectX 11 · HLSL

A custom real-time 3D rendering engine built from scratch, demonstrating low-level GPU pipeline programming, shader architecture, advanced lighting models, and procedural terrain generation.

Key Highlights:

• Full D3D11 pipeline initialization (device, swap chain, rasterizer state, depth-stencil)
• Shader compilation and management with an object-oriented hierarchy
• Phong lighting system with directional, point, and spot lights, fog, and materials
• Heightmap-based terrain generation with per-vertex smooth normal averaging
• Custom perspective camera with OpenGL-to-DirectX NDC conversion
• GPU resource lifecycle management (constant buffers, vertex/index buffers, COM cleanup)

---

Core Competencies Demonstrated

• Systems Engineering Modular architecture, memory/resource management, pipeline orchestration

• High-Performance Computing GPU compute shaders, multistage dispatch, data-parallel workloads, SIMD vectorization

• Graphics & Rendering Shader programming, 3D lighting models, GPU pipeline configuration, procedural geometry

• Performance Engineering Cache optimization, custom allocators, expression templates, algorithm tuning with measured results

• Machine Learning Deep learning architectures, training workflows, model optimization

---

Relevant Coursework

Core Computer Science

• CSC 400 — Discrete Structures for CS
• CSC 402 / 403 — Data Structures I & II
• CSC 406 / 407 — Systems I & II

Systems & Performance

• CSC 461 — Optimized C++
• CSC 562 — Optimized C++ Multithreading
• CSC 588 — Real-Time Multithreaded Architecture
• SE 456 — Architecture of Real-Time Systems
• CSC 486 — Real-Time Networking (In Progress)

Graphics & Game Engine Development

• GAM 425 — Applied 3D Geometry
• GAM 470 — Rendering / Graphics Programming
• GAM 475 / 575 — Real-Time Software Development I, II, & III
• GAM 576 — GPU Architecture

Machine Learning & Parallel Computing

• CSC 483 — Applied Deep Learning
• CSC 467 — CUDA Development (In Progress)

Independent Study

• Game Physics Project