Michael Liang

Developed and shipped Dart/Flutter features for the NetxTen application enabling real-time industrial gas leak detection. Designed and tuned signal processing methods to suppress noise and reduce false positives, improving reliability. Built and optimized C++/OpenCV pipelines for leak localization from raw sensor data to post-processing, and implemented fixes for edge cases in varying lighting and environmental conditions. Created evaluation workflows to measure performance and guide iterative improvements.

Built a chatbot using OpenAI, Gemini, Claude, Cohere, and DeepSeek APIs to generate immigration guidance from user inputs. Integrated OCR to parse documents automatically, reducing manual data entry. Implemented a RAG pipeline with LangChain and FAISS using 300+ past chats for context; set up evaluation using BLEU, ROUGE, and Sentence-BERT. Implemented safety checks to monitor hallucinations and enforce content rules.

Developed post-processing routines for object detection and image segmentation using OpenCV, Ultralytics YOLO, and PaddlePaddle. Improved model detection accuracy and edge alignment; reduced AI output errors by 50% and tripled Takeoff Boost product usage. Curated and preprocessed over 150,000 images to build a dataset; trained CNN and Vision Transformer models using Hugging Face. Evaluated model performance on GCS imagery with IOU, Dice, precision, recall, and F1 to guide training and threshold tuning. Deployed models as Vertex AI endpoints on GCP, managed Conda environments, and served inference APIs to 300 enterprise customers. Automated dataset refresh and model evaluation scripts to support continuous retraining of deployed vision models.