Chao Hsuan Ho

A full RAW→RGB image-signal-processing pipeline built two ways — a classical NumPy ISP versus a learned DnCNN denoiser — compared on the standard Kodak benchmark.

Constant-velocity Kalman filter, Hungarian IoU association, full track lifecycle, and a hand-written MOTA/MOTP/IDF1 suite validated against the reference motmetrics library.

DeepSORT (Kalman + cascade matching), a Transformer-based tracker, and a ResNet50 + channel-attention person re-identification model, implemented from the algorithms up.

A SIMD-vectorized sampling-based motion planner reimplementing the ICRA 2024 VAMP paper, with hierarchical sphere culling and SLEEF vector math for a 6-DOF arm.

Hands-on CUDA programming from vector-add through tiled GEMM and neural-network primitives, each with profiling and performance targets.

Data ingestion → RAG → LoRA/DPO fine-tuning → SageMaker deployment behind a FastAPI service; built on the LLM Engineer's Handbook architecture and extended with a test suite.