Tzu-Tsai Wei

Machine Learning Engineer with 4+ years of production experience, focused on closing the gap between research and deployment. Ships on-prem LLM/VLM and OCR systems that match or beat cloud APIs, and fine-tunes Transformer models end to end, from data pipeline to production inference. TOEIC 970/990.

Machine Learning Engineer at DeepQ (HTC) (2022-09 – Present)

Deep Learning Team, Healthcare AI

• Built an on-prem dual-OCR + VLM pipeline for bank voucher extraction: three local models combined for 96.9% vs. 93.1% for the best single model, matching then-SOTA cloud API (Gemini 3.5 Flash) at 2.5x the speed.
• Built a production OCR recognition model over 7+ versions on a rule-based synthetic-data and augmentation pipeline: 91% CER reduction on real healthcare samples, beating Azure and GCP cloud OCR, plus 360-degree rotation-invariant inference.
• Re-engineered Whisper Large's language-token space to add Hokkien, fine-tuning on 1,700+ hours of balanced trilingual media: Hokkien CER 54.3 to 28.5 (-48%), with 37% gains in English/Mandarin.
• Beat a raw-LLM baseline with LLM+CV, then a model-free CV pipeline on vaccine-form recognition: 72.4% to 87% to 100% per-image accuracy, GPU to CPU-only, 2.3s to 0.9s.
• Migrated the vision platform to a Swin Transformer backbone: 174% mAP gain in detection, 20-point Dice gain in segmentation (62.6 to 82.8), and segmentation training cut from 111h to 22h.
• Developed a patent-pending STT biasing algorithm for medical/gaming terms: 95.5% recall vs. Azure Phrase List's 19.8%, holding CER to 3.7% on CV19 (zh-TW) under a 1,019-term distractor stress test.
• Built an OCR-grounded LLM document digitization pipeline; on healthcare brochures, grounding cut CER 72% (0.075 to 0.021) and lifted BLEU 5 points by suppressing hallucination.
• Deployed and maintained an internal vLLM inference server serving DeepSeek-OCR2, Qwen and Dot-OCR variants over a shared API for team prototyping and evaluation.
• Built a data acquisition pipeline processing 2,300+ hours of trilingual broadcast media, growing Hokkien training data 18x to 1,883 hours vs. the 2023 open-source baseline (TAT-Vol1~2).

M.S. in Electrical Engineering (Machine Learning and Image Processing) – National Taipei University of Technology (2018 – 2021)

B.S. in Electrical Engineering – National Taipei University of Technology (2014 – 2018)