Work Experience

2023

2024

2025

ML Engineer Intern

Nuvoton

ML Engineer Intern

Nuvoton

May 2024 - August 2024

Skills:

Python, C++, Embedded Systems, Machine Learning, Audio DSP, Signal Processing, Data Augmentation, Model Optimization, Quantization, Confusion Matrices, CI/CD, Testing, Documentation, Problem Solving, Communication

Frameworks:

TensorFlow, TensorFlow Lite, scikit-learn, Keras, NumPy, Pandas, Matplotlib, GitHub Actions, Keil uVision
Accomplishments:

On-device sound event classification on MCU targets; baseline 49% → ~90%+ under defined test conditions
Built C++ test harness and Python evaluator; computed TP/TN/FP/FN, generated confusion matrices and hit-rate plots
Designed YAML-driven PipelineController for deterministic preprocessing: rename → segment → trim silence → normalize → augment → balance
Implemented selective normalization and gain jitter to preserve loudness cues and reduce false triggers
Tuned features: expanded MFCC 10→30 within MCU memory and timing limits; aligned window and stride with DSP constraints
Trained 30+ model variants (CNN, DS-CNN, LSTM); standardized configs for reproducible experiments
Exported minimal-operator TFLite models; quantized to int8 with ~60% size reduction and <2% accuracy delta
Engineered C++ KeywordSpottingSystem class and sliding-window InferenceFilter to stabilize predictions
Sized tensor_arena explicitly after profiling; prevented runtime OOM and fragmentation issues
Automated nightly CI jobs: preprocess → train → convert → flash → run tests → push CSVs and plots
Added distractor negatives to reduce false positives; established standard test conditions for volume and distance
Authored quick-start, pipeline guide, and troubleshooting notes; documented rationale for feature and threshold choices
Collaborated with firmware on linker scripts, DMA/I2S use, and memory maps; added guards for underflow/Na

Embedded ML Engineer

Nuvoton

Embedded ML Engineer

Nuvoton

August 2024 - Present

Skills:

Python, C++, Embedded Systems, Audio DSP, Signal Processing, Data Engineering, Model Optimization, Quantization, Memory Optimization, DMA, I2S, I2C, UI Development, Performance Testing, CI/CD, Documentation, Communication

Frameworks:

TensorFlow, TensorFlow Lite, Keras, NumPy, Pandas, Matplotlib, GitHub Actions, Keil uVision
Accomplishments:

Scope: productionizing edge AI stack, dataset expansion, robust MCU deployment with readable UI
Data ops at scale: YouTube and Pixabay scraping, mining and labeling, standardized file schema class_label_origin_name.wav, sidecar metadata and provenance index
Data loader redesign: AudioProcessor and DataLoader, shuffling and weighting, labeling modes, reduced loader from 525 to 407 lines, guarded edge cases
Scheduling and experimentation: unattended batch runs, background noise sweeps, 3D volume analyses, chosen plateaus documented in run book
Model ops: model_id metadata, auto config.txt emission, confusion matrix auto trim and transpose, consistent plots for comparison
Embedded architecture: KeywordSpottingSystem class, main.c reduced to a minimal loop, clean C and C++ boundaries
UI and display: DynamicText with bounded buffers and leak fixes, 100 ms refresh, collapsed repeats, field friendly readout
Memory and peripherals: display buffer 16 bit to 8 bit, RAM fit from about 648 kB down to near 510 kB, mapped interfaces, resolved PDMA conflict paths
I/O and buses: verified I2S and I2C plans, tuned priorities, added checks for NaNs and underflow to stabilize timing
Multi model deployment: command word spotting and sound event classification side by side, metadata printed at boot
Research and robustness: evaluated unknown handling strategies, class weighting, targeted negative sampling, failure log drove new data
Tooling and docs v2: eighteen pages across pipeline, APIs, training and deployment playbooks, integration diagrams
Cross team integration: reviews for linker flags, memory maps, DMA plans, serial formats; delivered demo ready builds