Biography
Xuhui (Daniel) Zhan is an Applied Scientist and Machine Learning Engineer working at the intersection of large-scale AI systems, multimodal learning, and ML infrastructure. He currently builds production machine learning systems at Treverse LLC, where he develops end-to-end AI solutions including a Go-based multilingual translation service deployed on AWS, an internal AI/ML platform supporting standardized training and CI/CD pipelines, an edge-first vision presort pipeline for high-speed barcode decoding, and recommendation models that generate personalized item suggestions for hundreds of thousands of users.
Previously at Vanderbilt University, Xuhui conducted research across multiple AI domains. At the AI Negotiation Lab, he led the development of an LLM pipeline that automatically codes negotiation transcripts, reducing annotation costs by over 99% while improving human–model agreement from 30% to 80%. His work has been presented at venues including the 2025 AI Negotiation Summit (Harvard/MIT). At the Network and Data Science Lab, he explored multimodal learning by integrating graph, vision, and language representations, including redesigning LLaVA-style architectures to improve training efficiency while maintaining competitive performance.
As a Professional Research Assistant, he applied computer vision to archaeological analysis, improving ancient mortar classification accuracy from 60% to 97% using Vision Transformers and designing a vision-only provenance algorithm that reduced similarity analysis time from months to minutes. Earlier in industry, he worked as an Algorithm Development Engineer, where he deployed computer vision storage-detection systems across approximately 30 warehouses and improved robotic localization pipelines for automated guided vehicles.
Xuhui holds a Master’s degree in Data Science from Vanderbilt University (GPA 3.98/4.0) and a Bachelor’s degree in Data Science with First Class Honours and Highest Distinction from Beijing Normal – Hong Kong Baptist University.
“Any sufficiently advanced technology is indistinguishable from magic.”
— Arthur C. Clarke, Clarke’s Third Law
- Natural Language Processing
- Computer Vision
- Graph Neural Network
- Multi-modality Model
- Recommender System
- Foundation Model
MSc in Data Science, 2023-2025
Vanderbilt University
BSc in Data Science, 2018-2022
Beijing Normal-Hong Kong Baptist University
BSc in Data Science, 2018-2022
Hong Kong Baptist University
Experience
Founding member of the Machine Learning Team
- Responsible for the full ML lifecycle across multiple projects and for building the core AI infrastructure
- Featured projects:
- Translation service implemented in Go and deployed on AWS
- Vision-based presort system for real-time barcode recognition
- Foundational AI infrastructure and recommendation systems built from scratch
Advisor: Prof. Ray Friedman
- Drive end‑to‑end LLM research for negotiation science—designing, experimenting, and optimising models.
- Built an LLM pipeline that auto-codes negotiation transcripts, cutting per-transcript cost from $5,000 to $3 (>99% savings) and boosting human–AI agreement from 30% to 80%.
- Developed a novel algorithm that injects population-level variation into LLM negotiator agents, yielding realistic and diverse subject pools for negotiation studies.
- Lead the full ML lifecycle—from data engineering through Web deployment—while collaborating closely with negotiation scholars and data-science peers.
- Work presented at the 2025 AINegotiationSummit (Harvard/MIT) and the International Association for Chinese Management Research annual meeting.
Advisor: Prof. Tyler Derr
- Developed a unified multimodal fusion framework integrating vision, graph, and language modalities:
- Vision–Language: Refactored LLaVA architecture, eliminating the projection pre-training stage, resulting in a 25% reduction in training time and halving data requirements while maintaining comparable performance across nine benchmarks.
- Graph–Text: Implemented a Temporal Graph Attention Network (TGAT)–LLM pipeline capable of generating high-quality textual attributes for social networks. Empirically evaluated performance on the Venmo Dataset, with extensive experiments on the Amazon Review Dataset forthcoming.
Graduate Teaching Fellowship
Responsible for organizing labs, creating quizzes, grading materials and hosting office hours.
Graduate level:
- DS 5620: Probability and Statistical Inference
- DS 5690: Gen AI in Theory and Practice
Undergrad level:
- DS 3100: Fundamentals of Data Science
Advisor: Prof. Markus Eberl
- Boosted ancient-mortar classification accuracy from 60% to 97% with Vision Transformers on 10M images.
- Invented a vision-only provenance algorithm, reducing similarity -analysis runtime from months to minutes; manuscript in preparation for Nature Methods.
Work as an algorithm engineer and take in charge of all machine learning applications and efficient data analysis.
Responsibilities include:
- Optimize the end localizer program and deploy on new products (ROS)
- Develop the vision algorithm for detecting whether two shelves are aligned (opencv, pytorch)
- Develop the end-to-end storage status detection system, implement the whole pipeline including data collection, data labeling, model training, model deployment, database construction, communication policies and api for updating status to central server (WCS) and display results on the front-end website with scaling abilities to aggregate and update results in real-time on constraint computation resources and deployment process (roboflow, django, redis, celery, pytorch, ONNX)