I focused on building and analyzing information related to companies in the primary market of Sichuan and Chongqing. My role primarily involved monitoring company news in areas such as business decisions, human resource management, and stock volatility.Key responsibilities included:- Data Scraping and Cleaning: I developed and implemented web scraping tools to gather relevant data from targeted websites. This included designing robust spiders capable of adapting to different site structures, filtering the collected data, and formatting it into concise, actionable insights.- Research and Implementation of Large Language Models: I worked on enhancing financial reasoning capabilities using advanced large-language models, including Meta's LLaMA-Chinese 2.0, LLaMA 3.0, Baidu’s ERNIE Bot, and ByteDance’s DOUBAO. I contributed to generating financial news summaries, integrating relevant data such as stock price charts, news reports, and company or government policy documents. Additionally, I optimized the pipeline for abstract generation, improving both efficiency and relevance.

I contributed to research aimed at developing an intelligent decision-making inference machine for asthma treatment plans. My specific focus areas included:- Logical Reasoning Architectures: I explored logical reasoning architectures grounded in expert domain knowledge, including Logical Neural Networks (LNN) and Logical Tensor Networks (LTN). This research enabled the generation of preliminary treatment plans via a systematic asthma treatment process.- Individual Treatment Effect Prediction: I conducted research on generative methods for predicting individual treatment effects, drawing upon NeurIPS and ICLR research such as GANITE and Dynamic DeepHit. This involved establishing a system for loading longitudinal medical data and adapting the GANITE architecture to facilitate individual data analysis. Additionally, I investigated the applicability of Dynamic DeepHit, initially designed for survival analysis, to medical risk assessment, considering the temporal nature of treatment effects.- Interpretability of Predictions: Leveraging the NeurIPS research SimplEX, I worked on enhancing the interpretability of the prediction process. This involved establishing a latent representation based on the similarity of individual patients' medical conditions, providing explanations for the prediction of new individual data.

- Audio Data Preparation and Feature Extraction: Manually labeling infant cry segments, merging annotations into labels, and applying audio signal processing techniques to extract filter bank features on a frame-by-frame basis.- Implementation of Supervised Deep Learning Models: Training supervised deep learning models such as Bidirectional LSTM, Transformer, and MobileNet V2 on the processed audio features and corresponding labels.- Research on Novel Unsupervised Methods: Investigating a new unsupervised method called Causal Representation Sparse Transition Clustering (CRSTC), based on causal temporal representation learning and clustering.- Designing Evaluation Metrics: Developing event-based F1 score and IOU metrics to evaluate models' ability to capture specific sound events within the time domain.