• Addressed ad cold-start limitations by training RQ-VAE and RQ-KMeans to convert multimodal ad embeddings (video, text, audio) into hierarchical Semantic IDs; improved token quality (>90% codebook utilization) and reduced ID collision rates by 20%+, enabling more accurate retrieval and ranking of new ads at scale with negligible computational overhead.

• Enhanced user interest modeling by optimizing the Deep Interest Network (DIN) to jointly capture both short-term and long-term behavioral sequences, incorporating cross-platform signals; boosted ad ranking performance with a +0.21% AUC gain, leading to more accurate ad targeting and improved user experience.

spatial-temporal understanding of VLM models

• Distributed fine-tuned LLaMA for function-calling ability with DeepSpeed, beated GPT-4's performance score in BFCL by 9.33%.

• Developed a LLM-based search agent with LangChain framework, enabled FastAPI calling, increasing informativeness by 10.31% and realism by 17.15% evaluated on a self-built dataset, based on an automated evaluation pipeline.

• Built a new dataset with 1k+ long context user-machine dialogs by web scraping, followed by data cleaning and editing.

• Maintained a Supabase vector database with BCEmbedding for RAG and long‑term conversation memory for a chatbot.

• Conduct an A/B test to evaluate the search agent’s ability to reduce cost, with a p-value of the t-test smaller than 0.01.

Designed prompts for anime characters and better RAG algorithm.

Applied langchain for auto-generating prompts and evaluations.

Training and generating songs for publicity.