Fine-tuned multi-lingual political and toxicity detection models (including mt5, FastText, and RoBERTa) by modifying neural network layers and utilizing LoRA and PEFT techniques, achieving a 0.97 recall rate against manual labels;Innovated a multi-model voting mechanism and a transfer learning approach using mixed training on eight minority languages, along with Chinese and English, achieving a 15% accuracy improvement compared to monolingual models;Developed a novel recursive sampling technique based on autoregressive weighted averages of sensitive word density, dynamically updating sampling ratios across different data sources, increasing hard case representation.

Developed a standardized data analysis and machine learning platform for analyzing sales records across different business units. The platform attributed business performance to specific driving factors and improved data quality.Built a Q&A system using Retrieval Augmented Generation (RAG) techniques. Fine-tuned large language models to generate insights and improvement suggestions, providing decision-making support for senior managers.Applied feature engineering to create multidimensional user portraits and trained a recommendation system using collaborative filtering and Restricted Boltzmann Machines, enabling precise insurance product recommendations.Conducted time series analysis based on sales data to provide accurate data support for the agents and to make plans.

Processes time series data based on Bi-RNN and LSTM to capture complex trends in Ethereum price changes.Integrates GRU technology and Dropout layer regularization to ensure generalization and accuracy of the model. Introduced CNN, attention mechanism, and autoencoder to further enhance the model's ability to capture price trends.The error rate is less than 0.05% in the test for predicting the next day's closing price for 64 consecutive days.

Developed a model for adding invisible watermarks to text generated by LLMs to improve its traceability.Customized the watermark detection algorithms based on specific watermark-adding models, increasing the detection accuracy rate to 99.3% and reducing the false positive rate to 1.6%, demonstrating high accuracy.Designed a variety of genAI-based methods to tamper with the watermark, proving the robustness of the model.Added a perplexity monitoring module based on Transformer to make the outputs more like human-written ones.

Developed an R language package integrating regression and evaluation models such as Lasso, Ridge, and KNN to fit the financial data of listed companies and predict development trends.Finished four research reports on listed companies, analyzed investment strategies, and participated in planning and organizing three investor education activities. Analyzed clients' financial situations and participated in 2 wealth planning projects to provide clients with wealth planning advice to meet the financial target of a 10% annualized return.