Key Skills

Software

Clickworker

Data Annotation Tech

Google Cloud Vertex AI

Labelbox

Lionbridge

OneForma

Remotasks

Scale AI

Telus

Top Subject Matter

No subject matter listed

Top Data Types

Image

Text

Video

Top Task Types

Action Recognition

Classification

Data Collection

Evaluation Rating

RLHF

Freelancer Overview

I have extensive experience in the end-to-end AI training data lifecycle, specializing in the creation of high-quality datasets for both Large Language Models (LLMs) and Computer Vision systems. My expertise includes a deep mastery of Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), where I have consistently delivered precise annotations for complex linguistic nuances and visual patterns. I am highly proficient in industry-standard tools such as Labelbox, Scale AI, and Prodigy, and I pride myself on maintaining exceptional Inter-Rater Reliability (IRR) scores. By focusing on rigorous quality assurance and the identification of rare edge cases, I have directly contributed to significant improvements in model accuracy and the reduction of algorithmic bias in production-ready environments.

Labeling Experience

Multilingual Speech Recognition & Acoustic Emotion Mapping

TelusAudioClassificationEmotion Recognition

I worked on a large-scale data annotation project designed to improve the Natural Language Understanding (NLU) of a global virtual assistant. My primary responsibility was the verbatim transcription of varied acoustic data, including noisy environments, cross-talk, and heavy regional accents. I performed Phonetic Labeling and tagged Acoustic Events (e.g., background noise, non-speech vocalizations) to help the model distinguish between user commands and environmental interference. Additionally, I contributed to a specialized Emotion Recognition layer, where I categorized the speaker's sentiment (e.g., frustrated, satisfied, neutral) and urgency levels. This data was used to refine the model's ability to adjust its tone of voice in response to user mood. By adhering to strict orthographic guidelines and maintaining a high throughput, I helped decrease the Word Error Rate (WER) for the client's localized speech models.

I worked on a large-scale data annotation project designed to improve the Natural Language Understanding (NLU) of a global virtual assistant. My primary responsibility was the verbatim transcription of varied acoustic data, including noisy environments, cross-talk, and heavy regional accents. I performed Phonetic Labeling and tagged Acoustic Events (e.g., background noise, non-speech vocalizations) to help the model distinguish between user commands and environmental interference. Additionally, I contributed to a specialized Emotion Recognition layer, where I categorized the speaker's sentiment (e.g., frustrated, satisfied, neutral) and urgency levels. This data was used to refine the model's ability to adjust its tone of voice in response to user mood. By adhering to strict orthographic guidelines and maintaining a high throughput, I helped decrease the Word Error Rate (WER) for the client's localized speech models.

2024

High-Precision Semantic Segmentation & Object Detection for Autonomous Systems

LabelboxImageBounding BoxPoint Key Point

In this project, I performed pixel-level Semantic Segmentation on complex urban street scenes to train perception models for autonomous vehicles. My work involved masking diverse classes, including driveable surfaces, pedestrians, traffic signage, and environmental obstacles. I maintained high consistency across large datasets, ensuring that boundaries were precise to within 1-2 pixels to prevent model "bleeding" between adjacent objects. In addition to segmentation, I executed Bounding Box annotation for object detection tasks, specifically focusing on identifying occluded (partially hidden) objects and varying lighting conditions (night-time/glare). I also performed quality audits on peer-labeled data to maintain a strict 98% accuracy threshold, providing feedback to improve overall dataset reliability.

In this project, I performed pixel-level Semantic Segmentation on complex urban street scenes to train perception models for autonomous vehicles. My work involved masking diverse classes, including driveable surfaces, pedestrians, traffic signage, and environmental obstacles. I maintained high consistency across large datasets, ensuring that boundaries were precise to within 1-2 pixels to prevent model "bleeding" between adjacent objects. In addition to segmentation, I executed Bounding Box annotation for object detection tasks, specifically focusing on identifying occluded (partially hidden) objects and varying lighting conditions (night-time/glare). I also performed quality audits on peer-labeled data to maintain a strict 98% accuracy threshold, providing feedback to improve overall dataset reliability.

2023

Advanced RLHF & SFT for Multi-Step Reasoning and Code Generation

Scale AITextRLHFFine Tuning

I served as a Tier-3 Subject Matter Expert (SME) focused on aligning frontier Large Language Models for better logical reasoning and programming accuracy. This project involved a dual-track workflow: Reinforcement Learning from Human Feedback (RLHF) and Supervised Fine-Tuning (SFT). I evaluated thousands of model-generated responses, ranking them based on a multi-dimensional rubric that prioritized code efficiency, security (detecting vulnerabilities), and conversational helpfulness. A significant portion of my work involved "Red Teaming" prompts to identify edge cases where the model might produce hallucinations or non-functional code.

I served as a Tier-3 Subject Matter Expert (SME) focused on aligning frontier Large Language Models for better logical reasoning and programming accuracy. This project involved a dual-track workflow: Reinforcement Learning from Human Feedback (RLHF) and Supervised Fine-Tuning (SFT). I evaluated thousands of model-generated responses, ranking them based on a multi-dimensional rubric that prioritized code efficiency, security (detecting vulnerabilities), and conversational helpfulness. A significant portion of my work involved "Red Teaming" prompts to identify edge cases where the model might produce hallucinations or non-functional code.

2023

Education

Work History