Jun 2025 – Sep 2025 · 3 mos

May 2025 – Present · 1 yr · Remote

Jun 2024 – May 2025 · 11 mos · Nashville, Tennessee, United States · Hybrid

• Advisor: Prof. Tyler Derr (Network and Data Science Lab)
• TL;DR: Worked on various research projects across multimodal models, foundation models, graph neural networks and recommender systems, aimed at developing an innovative unified fusion framework to potentially replace LLaVA (and other soft prompt/token based multimodal frameworks), achieving improved performance and efficiency when integrating multiple modalities.
• Proposed a unified fusion framework that seamlessly integrates vision, graph, and language modalities.
• Vision–language: adapted fusion framework on the LLaVA code‑base, eliminated the alignment pre‑training stage; relied on fine‑tuning alone cuts total training time ≈ 25 % and halves data needs while matching SOTA performance on nine public benchmarks. (See the project page for Inverse-LLaVA at https://inverse-llava.github.io)
• Graph–text: built an advanced TGAT‑LLM pipeline leveraging proposed fusion framework to generate rich textual attributes for social‑network data; full experimental runs on Venmo and Amazon Reviews are in progress, with encouraging preliminary results.

Oct 2023 – May 2025 · 1 yr 7 mos · Nashville, Tennessee, United States · Hybrid

• Advisor: Prof. Ray Friedman (AI Negotiation Lab) | Paid Position
• TL;DR: Prompt engineering, Retrieval-Augmented Generation (RAG), Large Language Model (LLM) research and applications—including fine-tuning, quantization, and curated data pipelines—as well as interdisciplinary collaboration and web development.
• Drive end‑to‑end LLM research for negotiation related applications—designing, experimenting, and optimising models.
• Auto‑coding project: cut per‑transcript cost from $5 k to $3 (>99 % savings) through an LLM pipeline refined via active‑learning loops.
• AI subject‑pool project: developed a novel algorithm that injects population‑level variation into LLM agents, creating realistic, diverse negotiator profiles.
• Collaborate closely with negotiation faculty and data‑science peers; lead model ops and mentor junior coders.
• Work presented at the 2025 AI Negotiation Summit (Harvard/MIT) and the IACMR annual meeting.

Oct 2023 – Jan 2025 · 1 yr 3 mos · Nashville, Tennessee, United States · Hybrid

• Advisor: Prof. Markus Eberl | Paid Position
• TL;DR: Computer vision-related research and applications, including handling imbalanced datasets, designing evaluation metrics, and implementing MLOps practices.
• SOTA classification pipeline: built an end‑to‑end Vision Transformer workflow that outperforms legacy methods.
• Provenance by vision: devised a sample‑similarity algorithm that traces mortar movement routes by comparing visual features alone, offering a non‑destructive alternative to chemical assays.

Aug 2023 – May 2025 · 1 yr 9 mos · Nashville, Tennessee, United States · Hybrid

• Position funded by the Vanderbilt Graduate Teaching Fellowship
• Led labs, designed quizzes, graded coursework, and held weekly office hours for:
• DS 5620 Probability & Statistical Inference (Grad), DS 5690 Generative AI: Theory & Practice (Grad) and DS 3100 Fundamentals of Data Science (Undergrad)
• Served as the dedicated statistics tutor for Master of Finance students in Econometrics I.

Jun 2022 – Apr 2023 · 10 mos · Suzhou, Jiangsu, China · On-site

• TL;DR: End-to-end, cost-sensitive, and scalable machine learning pipelines covering design, development, and deployment (OpenCV, YOLO, PyTorch → Django, ONNX, Async, Celery, Redis, CPython), as well as Robot Operating System (ROS) software development and deployment in C++.
• Optimized and deployed high‑precision localization algorithms for Automated Guided Vehicle (AGV) robots, improving positional accuracy and enabling rollout to new product lines.
• Built an end‑to‑end vision storage‑detection system (YOLO · OpenCV · PyTorch→ONNX · Django) that integrates with the central server for large‑scale, multi‑AGV coordination and visualization; standardized hardware and deployment playbooks for rapid installs in  more than  30 warehouses.