Projects

Situation

Target embedded hardware (M5Stack V2unit) had severely constrained compute and memory, demanding a highly optimized model and deployment workflow for real-time environmental sound classification.

Task

Architect and deliver a complete, reproducible model training-to-edge-deployment pipeline targeting an embedded Linux device.

Action

Engineered a Mel spectrogram feature extraction pipeline on the ESC-50 dataset with audio augmentation; trained a CNN classifier in PyTorch using 5-fold cross-validation and conducted Adam vs. SGD optimizer ablation; converted the trained model via PyTorch → ONNX → NCNN and cross-compiled the inference runtime for the M5Stack V2unit via SSH.

Result

Delivered a fully operational edge inference system with stable real-time performance on embedded hardware, validating the complete PyTorch → ONNX → NCNN cross-compilation workflow.

Situation

Traditional manual attendance in classroom settings was time-consuming and error-prone, with no automated hardware-level solution available at low cost.

Task

Architect a low-cost, highly reliable multi-sensor fusion system to fully automate classroom attendance tracking.

Action

Architected a multi-sensor pipeline integrating an AS608 optical fingerprint module, a pyroelectric infrared (PIR) presence sensor, and a DS18B20 digital temperature sensor; authored modular C/C++ firmware to maximize reusability and portability, significantly reducing BOM cost and system complexity.

Result

Co-invented the system architecture and was granted a National Utility Model Patent by CNIPA, confirming the solution's technical novelty and industrial applicability.