cv

AI Agent Development Intern, Digital Intelligence Support Center

China Mobile Communications Group Jiangsu Co., Ltd.

Built an internal AI-agent workflow system to automate quarterly/annual operational report generation; integrated LLM-driven extraction, summarization, and templated drafting with cross-department stakeholders.

• Delivered 3 reusable workflows covering 10 report sections to standardize formatting and speed iteration
• Prototype adopted internally; recognized for independence, communication, and execution by the AI Group

Research Assistant (Advisor: Ming Li)

Speech and Multimodal Intelligent Information Processing (SMIIP) Lab, Duke Kunshan University

Research on real-time whisper-to-normal speech conversion using neural audio codecs with Transformer-based mapping; focus on robust training under misalignment and evaluation on real whisper datasets.

• Built codec + Transformer speaker-conditioned mapping with DTW-based latent alignment and multi-objective waveform/latent losses
• Benchmarking on wTIMIT and AISHELL-6-Whisper; evaluating with WER/CER, DNSMOS, SI-SDR, and MCD
• First-author manuscript in preparation for ICME 2026

Summer Research Scholar (Advisor: Paul Weng)

Duke Kunshan University

Built and evaluated a Retrieval-Augmented Generation (RAG) pipeline for question answering using Llama-3 8B; improved retrieval quality via indexing, embedding tuning, and filtering.

• Implemented hierarchical indexing and embedding fine-tuning; ran 75 hyperparameter configurations
• Added relevancy check (cosine similarity threshold 0.6) to filter irrelevant queries and reduce cost
• Achieved +21.4% Hit Rate and +12.9% Context Recall; improved faithfulness (+8.3%) and context precision (+9.1%)

Research Assistant (Advisor: Huansheng Cao)

Duke Kunshan University

Multi-omics analysis of cyanobacterial bloom suppression under extreme summer heatwaves; large-scale pipelines on computing clusters and biological interpretation of heat-driven dynamics.

• Processed metagenomic, metatranscriptomic, and metabolomic datasets (QC, taxonomic profiling, functional annotation, integration)
• Quantified species-level shifts and thermal thresholds suppressing Microcystis blooms
• Co-first author manuscript submitted to Nature Communications (under consideration)

Senior Program Assistant (Student Worker)

Innovation and Entrepreneurship Initiative (InE), Duke Kunshan University

Coordinated and executed entrepreneurship programming and incubator events; planned operations and on-site logistics to improve engagement and event quality.

• Supported Entrepreneur Speaker Series (ESS) and U-Corp Lab, averaging ~60 attendees per event
• Helped run Innovation Incubator (Dii) events serving 100+ incubator teams (Innovation Challenge, Chinese College Student Innovation & Entrepreneurship Program)

B.S. (Dual Degree)

Duke Kunshan University (DKU) & Duke University

Data Science; Interdisciplinary Studies (Data Science)

Whisper-to-Normal Speech Conversion (Real-time)

A real-time whisper-to-normal speech conversion system using neural audio codecs and a Transformer-based, speaker-conditioned mapping with DTW-based latent alignment and multi-objective losses.

• DTW-based latent alignment to handle misaligned real whisper datasets
• Evaluation: WER/CER, DNSMOS, SI-SDR, MCD; ablations on alignment and multi-objective losses
• Target datasets: wTIMIT, AISHELL-6-Whisper

Multi-omics Analysis of Bloom Suppression under Heatwaves

End-to-end processing and integration of metagenomic, metatranscriptomic, and metabolomic data to study temperature-driven microbiome dynamics and thermal thresholds affecting Microcystis blooms.

• Species-level dynamics + thermal threshold quantification
• Strain-level genomic analysis for M. aeruginosa and associated phages

RAG Question Answering with Llama-3 8B

A retrieval-augmented question answering pipeline with hierarchical indexing, embedding fine-tuning, and relevancy filtering to improve retrieval and reduce unnecessary compute.

• 75-config hyperparameter sweep; best model improved Hit Rate (+21.4%) and Context Recall (+12.9%)
• Relevancy check via cosine similarity thresholding (0.6)