Paul O.

Key Skills

Software

Mindrift

Micro1

OneForma

Remotasks

Top Subject Matter

AI Model Evaluation (RLHF)

Data Quality & Annotation

nstruction Following & Creative Writing

Top Data Types

Image

Video

Text

Top Task Types

Text Generation

RLHF

Evaluation/Rating

Data Collection

Object Detection

Transcription

Freelancer Overview

My experience is rooted in the high-stakes environment of AI model alignment, where I function as both an evaluator and an architect of complex instructions. Having worked with platforms like Outlier and Handshake, I have developed a specialized skill set in Reinforcement Learning from Human Feedback (RLHF). My day-to-day involves meticulously auditing model outputs for factual accuracy, logical soundness, and stylistic precision, ensuring that the AI’s responses are not just grammatically correct, but genuinely useful and safe. What sets my work apart is my ability to thrive in the "human-in-the-loop" ecosystem. I excel at deconstructing ambiguous or multi-layered user prompts into clear, executable steps, which is critical for reducing hallucinations and improving reasoning capabilities. By identifying patterns in model failures and providing granular, actionable feedback, I bridge the gap between raw data and high-performance AI, ensuring that models reliably meet the rigorous quality standards required for professional, technical, and creative applications.

Labeling Experience

AI Training Specialist & Data Annotation Specialist

ImageRLHF

As an AI Training Specialist, I contributed to high-volume model alignment projects focused on Factuality & Grounding. My core responsibilities included executing RLHF workflows by evaluating, ranking, and refining model responses to ensure factual accuracy, safety, and adherence to complex, evolving stylistic guidelines. I managed many of individual data points, consistently maintaining a quality score of 100%. My process involved identifying and correcting subtle logical lapses and hallucinations, providing granular feedback to improve model performance, and strictly following rubric-based instructions to ensure outputs were highly calibrated for end-user utility

2025 - Present

Education

Work History