📚 Efficient RAG(Retrieval-Augmented Generation) system and framework development 🏷️Unstructured Data Chunking and Retrieval (Multimodal Modeling Solutions)

https://drive.google.com/file/d/14MuDngvDFcidYnW_a-Tob-IaYTTBNZ4p/view?usp=sharing - Assessed and integrated Generative AI, including LLM Agents(LangChain) and Visual-language Multimodal Models, LLM Customization(RLHF, PEFT, Prompt, etl...), into electric vehicle projects, focusing on pioneering applications and surpassing traditional technology barriers.(https://hackmd.io/25Qd-n1NQKOqBUlsmm7NNw) - Successfully custom deployment of YOLO-Pose on edge devices using Deepstream, achieving real-time AI inference. Gained notable recognition on GitHub(50 stars + fork ).( (GitHub: https://github.com/YunghuiHsu/deepstream-yolo-pose) - First to build, improve, and monitor RTSP streaming pipelines and latency using the Gstreamer Python API to ensure seamless HPC data transfer for electric vehicles.(https://hackmd.io/r9JWCizqSSenrggXATKn-w) Integrated advanced computer vision algorithms, such as YOLOv7, enhancing edge computing performance. (Details: https://hackmd.io/@YungHuiHsu/BJ7fpQyps) - Boosted model inference speed by 3-6x on edge devices through techniques like "reparameterization" and "quantization", paired with TensorRT and DeepStream(Details: https://hackmd.io/@YungHuiHsu/HybM-QvEn) - Developed a Self-Supervised Deep Learning algorithm to extract meaningful features from millions of unlabeled time series data, specifically for anomaly detection in driving behavior. (GitHub: https://github.com/YunghuiHsu/IEM_DORA)

https://docs.google.com/document/d/1gxWLJp6wZB8CrES8KW6amFFW2Z1ekpAoMbvB9lMM84s/edit 1. Developed task-specific segmentation model from scratch - Integrated unsupervised & supervised learning. - Enhanced model perf.: Achieved IoU <0.2 to 0.93, boosted background seg. accuracy to 99.3% on 15k images. - Implemented customized loss function, hierarchical sampling & data augmentation 2. Led a self-supervised learning project focused on extracting features from unstructured data(images), strengthening representation learning capabilities. - Introduced and fine-tuned a SOTA model (MAE with ViT), boosting accuracy from 75.9% (CNN) to 95.7% (ViT). - Implemented explainable AI (XAI) methods for Transformer reliability. - Implemented MAE sparse coding to restrict embedding distribution, boosting sparsity from 43% to 84% and enhancing analyzability.

- Use statistical methods to analyze complex data sets and provide insight into conservation strategies. - Translated data into concise visual reports, facilitating stakeholder understanding for informed decisions. - Served as a central communicator in program management, bridging gaps between various entities, including government departments and civil society. - Rapidly mastered Python to automate data processing, enhancing efficiency and expediting results. - Consistently adapted to and learned new tools, furthering operational excellence and value to the company.