1.Improved query–merchant relevance by +2.13% Accuracy and +2.07% Recall through multi-stage fine-tuning of cross-encoder mBERT with feature augmentation, leading to +1.02% UV_CTR and +0.83% GMV in online A/B tests while resolving 83% of bad cases.

2.Expanded training data by 300K+ samples with labeling accuracy from 70%+ to 90%+ via user behavior–based sampling (view, click, order) and LLM-enhanced labeling.

3.Boosted benchmark metrics across full-intent queries by implementing feature augmentation on query and merchant sides through multi-stage fine- tuning by designing multiple groups of experiments of user behavior-based training samples and original ones.

4.Streamlined offline pipeline by delivering end-to-end relevance modeling, including sample generation, LLM labeling, training, inference, and caching.

1.Built a To-B chatbot integrating with a RAG pipeline including data pre-processing, text splitter, text embedding model, vector

database, and large language model (LLM) in Python to help business customers quickly learn details of robot products.

2.Converted HTML contents of text and tables into text format usingBeautifulSoup4 and stored all data including id, text, url, title, and

labels into MongoDB as the database.

3.Used Cross-segment BERT and text-splitter in LangChain to split the Chinese and English text into 30,000 chunks of different

lengths semantically.

4.Generated dense and sparse vector embeddings from text chunks using text-embedding-v2 from Ali DashScope and stored them

into Ali DashVector with multiple attributes for content-based and keyword-based recall.

5.Used TF-IDF algorithm to extract robot name keywords from user queries to help filter out intented contents of specific robots.

6.Fine-tuned Qwen-plus (LLM) by using proper prompting and top-k reranked recommended documents including text, title, and url from vector database based on cosine similarity and dot product to generate responses to users.

7.Deployed the RAG-based chatbot to Docker image and official company website and community using Flask.

1.Designed automated pipeline and evaluated models for automatic speech recognition (ASR) correction based on the WikiHow dataset, generating new knowledge base, and improving the speech-to-text performance of task-oriented robots by 30%.

2.Addressed challenges in question type recognition, prerocessed conversational text datasets, and trainedquestion classifiers to improve the robot's NLU.

3.Improved accuracy of intent recognition to enhance the ability to analyze user instructions.

4.Developed multi-model user interfaces for Echo device based on Alexa Presentation Language, incorporating voice, text, and video

interactions to enhance user satisfaction by 20%.

5.Published a research paper “TACO 2.0: A Task-Oriented Dialogue System with Mixed Initiatives and Multi-Modal Interaction“ on Amazon Science.

Analyzed specific audio datasets and generated corrupted audio files and transcriptions. Searched for and analyzed the state-of-the-art ASR models. Created a pipeline including Text-to-speech model and Speech-to-text model. Made improvements on ASR error correction model.

1. Developed and deployed a convolutional neural network (CNN) for malware classification, achieving 98.7% accuracy on a dataset of 200,000+ binary samples by converting binaries into grayscale images for feature extraction.

2. Optimized the CNN architecture with advanced layers and hyperparameters, implementing data augmentation strategies to enhance model robustness against new malware variants.

3. Integrated the model into a real-time cloud-based malware detection service, using Docker and Kubernetes for scalable deployment, processing thousands of files per minute.

4. Collaborated with the engineering team to integrate the system into the cybersecurity platform, conducting performance evaluations and ensuring seamless API development and deployment.