Linh P.

Key Skills

Software

Don't disclose

Top Subject Matter

Medical VR training simulation

LLM fine-tuning for financial Q&A

Top Data Types

3D Sensor

Text

Top Task Types

Question Answering

Prompt + Response Writing (SFT)

Data Collection

RLHF

Freelancer Overview

I have experience working on AI training and evaluation tasks involving long-horizon reasoning, instruction following, prompt engineering, and environment-based task validation for remote AI data operations projects. My work has included designing and testing complex multi-step tasks across productivity and workflow environments, evaluating model behavior against detailed guidelines, identifying failure cases, and refining prompts to improve reliability and reasoning quality. I have also worked on data structuring, annotation consistency, QA feedback incorporation, and edge-case analysis, particularly in scenarios where models must maintain context, follow constraints, and complete realistic user workflows accurately. Through these projects, I developed strong attention to detail, systematic evaluation skills, and the ability to quickly adapt to new tools and task requirements. Beyond commercial AI training tasks, my research background strengthens my ability to work with AI datasets critically and analytically. As a Computation and Design student at Duke Kunshan University, I have worked on interpretable AI and immersive technology research projects involving behavioral data collection, sensor-driven systems, and human-centered evaluation pipelines. I have experience organizing structured datasets, working with multimodal data, evaluating model outputs for consistency and interpretability, and building workflows that require both technical precision and user-centered thinking. My combination of AI evaluation experience, research-oriented analytical skills, and strong adaptability allows me to contribute effectively to high-quality training data and model improvement workflows.

Labeling Experience

Undergraduate Researcher, Trust and Reliability in Sketch-Based AI Systems for VR

3D Sensor

Designed and built a sketch-based AI system for behavior-grounded personality assessment in VR. The project required capturing user sketch and motion-derived behavioral signals and transforming them into model inputs for generating natural-language personality feedback. It also involved designing controlled user studies using validated trust scales to evaluate feedback stability and how users calibrate trust. • Built a minimalist VR sketching environment with controller trajectory tracking and stroke-feature extraction (e.g., speed, smoothness, pauses, symmetry, density). • Encoded behavioral/visual sketch features and integrated with the Gemini API to generate natural-language personality feedback. • Designed a controlled user study using validated trust scales and feedback-stability manipulations. • Analyzed how users calibrate trust in sketch-based AI systems based on generated feedback outcomes.

2025 - Present

Scientist Research Intern, GRIPS Program (Fudan University)

Prompt + Response Writing (SFT)

Developed and evaluated an instruction-tuned language model for a domain-specific financial question answering task. The work included designing a LoRA-based fine-tuning pipeline and using a supervised fine-tuning trainer to learn from a financial Q&A dataset. Performance improvements were quantified using prompt-based evaluations of reasoning completeness and financial terminology correctness. • Conducted literature review on Transformer attention, pretraining vs. fine-tuning, and autoregressive decoding. • Designed and implemented a LoRA fine-tuning pipeline for Qwen2.5-1.5B-Instruct using Hugging Face SFTTrainer. • Used a domain-specific financial Q&A dataset for training/instruction tuning. • Achieved ~25% improvement in answer accuracy via prompt-based evaluation of reasoning completeness and terminology correctness.

2025 - 2025

Research Assistant, iRow Exergame (Human-Interactive Intelligence Lab)

Don't discloseQuestion Answering

Worked on an AI-driven VR exergame research effort that required creating and evaluating user-facing adaptive generative content. The work involved interpreting how users understand AI-generated adaptive prompts/content and assessing related engagement and performance outcomes in VR exercise settings. This included coordinating how biometric sensing outputs were mapped to adaptive logic that altered the generative environment during sessions. • Designed and conducted a pilot usability study with mixed-methods instrumentation for engagement, comfort, and safety. • Coordinated with technical contributors to align biometric sensing, adaptive logic, and gameplay instrumentation. • Analyzed user interpretation of AI-generated adaptive content, focusing on transparency, immersion, and cognitive load design tensions. • Supported research iteration cycles based on user study findings and feedback from the team.

2025 - 2025

Education

Work History