AI engineer with 8 years of production machine learning across speech, NLP, and computer vision — currently shipping real-time, on-device(mobile/MCU) audio-related model at Savitech. Equally an educator: 11+ technical talks at Taiwan AI Academy and corporate training for Qualcomm, Taiwan Mobile, and Innolux.
工程與教學並重，擅長把模型從論文帶到產品。
EXPERIENCE 工作經歷
Savitech 盛微先進科技 — Senior AI Engineer
Jul 2021 – PresentBuilt the end-to-end personalized HRTF (spatial audio) system; ship real-time, on-device audio models to mobile and MCU targets.
AI Foundation 人工智慧科技基金會 — Instructor
Apr 2021 – Apr 2022Transformer training for engineering staff, whiteboard derivations.
AI4kids — AI Engineer
Mar 2021 – May 2021AI Foundation — AI Engineer
Sep 2020 – Mar 2021Taiwan AI Academy 台灣人工智慧學校 — AI Engineer
Jul 2018 – Sep 2020Three industry projects; 11 technical talks on deep learning and NLP.
SELECTED AI PROJECTS 精選專案
Music Separation — reproduced paper to 10.75 dB vocal cSDR on MUSDB18 (SOTA for on-device-scale models), then ported it to C for mobile and MCU deployment(real-time).
Voice Enhancement — Built the industry's first-in-house 8 kHz → 48 kHz speech bandwidth-extension model, validated by the acoustics team, and ported it to C for deployment on mobile devices and MCUs. (Library size 1.6MB)
Noise Suppression (ENC) — convert model to C and real-time on-MCU noise suppression ; CNN encoder–decoder, quantization, distillation (Library size 200KB+).
Personal Spatial Audio HRTF system — Built an end-to-end system that creates personalized 3D spatial audio (HRTF) from a single photo, from on-device perception models to an automated server pipeline.
Image2Mesh — Reworked MeshTransformer (full-body → face/ear) for single-image 3D mesh reconstruction; built the paired image-to-mesh dataset and trained the model.
Mesh2HRTF — Generate 3000+ PHRTF.
Face/Ear Object Landmark Detection — Built the complete pipeline (data collection → annotation → Yolov7 → Android/iOS), delivering real-time Mobile inference.
棒球賽事智能剪輯 — CV auto-extracts pitch/hit clips (InceptionResNet, Bézier curves).
2D-code defect detection — real-time production-line CV inspection.
棒球賽事紀錄轉換數據 — NLP converts CPBL records into structured data.
TEACHING 教學經歷
協助公司導入 Copilot / Claude — training engineer teams on skills, agents, and workflows — cutting time on routine engineering tasks by 75%.
Taiwan AI Academy — 11 talks: Deep Learning, Back-propagation, RNN, Seq2seq, Attention, Transformer, Transformer-XL, XLNet, BERT, GPT-3, PCA.
Corporate — Qualcomm 高通, Taiwan Mobile 台灣大哥大, Innolux 群創光電, Powertech 力成科技, 微程式資訊, 大力卜企業.
Community — Taichung.py, public Transformer Workshop, 逢甲 AI Lab instructor.