Mary Mandefu

Key Skills

Software

Appen

Labelbox

Mercor

OneForma

Remotasks

Telus

Other

Scale AI

Internal/Proprietary Tooling

Top Subject Matter

No subject matter listed

Top Data Types

Audio

Image

Text

Top Task Types

Classification

Evaluation/Rating

Fine-tuning

RLHF

Text Summarization

Freelancer Overview

I have built strong expertise in AI data labeling, model evaluation, and training data quality assurance through my work with Outlier, Invisible Technologies, RWS, TELUS, and Transcription Puppy. My experience includes evaluating AI-generated outputs, refining prompts, developing test cases, and contributing to RLHF pipelines, as well as working extensively with structured data formats such as JSON and YAML. I am skilled at identifying inconsistencies, logical gaps, and missing assumptions, and I apply detailed guidelines to ensure accuracy, clarity, and real-world alignment across a wide range of annotation and evaluation tasks. What sets me apart is my ability to combine deep analytical reasoning with strong documentation and structured evaluation skills. My background in geospatial analysis and planning has strengthened my ability to interpret complex datasets, design logical frameworks, and produce clear, concise assessments. I am comfortable reviewing and refining scenarios, defining gold-standard behaviours, and annotating reasoning paths to support high-quality model training. With a strong attention to detail, excellent written communication, and disciplined remote-work habits, I consistently deliver accurate, high-integrity training data that enhances AI performance.

Labeling Experience

Advanced Data Trainer

Internal Proprietary ToolingImageEvaluation Rating

- Evaluating the preservation of subject identities in two images - Evaluating any artifacts, limb disfiguration, and/or other logical issues in the images (either Source or Output) - Evaluating logical issues: If the source and target image are photorealistic, i.e. do they look realistic with no factual errors? e.g., having a woman's face on a man’s body etc. - Evaluating the expressions in the Source Image and Target Image

- Evaluating the preservation of subject identities in two images - Evaluating any artifacts, limb disfiguration, and/or other logical issues in the images (either Source or Output) - Evaluating logical issues: If the source and target image are photorealistic, i.e. do they look realistic with no factual errors? e.g., having a woman's face on a man’s body etc. - Evaluating the expressions in the Source Image and Target Image

2025

Data Annotator

Internal Proprietary ToolingImageData Collection

- Search for Instagram Reels, pause at specific frames, and tag entities (people, clothing, products, locations, etc.) with reference photos. - Search Instagram images and tag entities. - Draw masks on images and generate 2 variants with different seeds, write a reverse prompt, then select your preferred result. - Create a conversational flow that alternates between user messages (image+text or text-only) and assistant image responses.

- Search for Instagram Reels, pause at specific frames, and tag entities (people, clothing, products, locations, etc.) with reference photos. - Search Instagram images and tag entities. - Draw masks on images and generate 2 variants with different seeds, write a reverse prompt, then select your preferred result. - Create a conversational flow that alternates between user messages (image+text or text-only) and assistant image responses.

2025

AI Data Trainer

Scale AITextClassificationText Summarization

- Rate two model generated responses separately across various dimensions. Provide preference ranking scores. Write a justification for the response preference ranking and explain the specifics and logic behind the decision - Evaluate two AI responses to prompts designed to help provide services to business users - Create data on how AI models should handle user interactions that deal with sensitive topics or harmful themes. - Write prompts requesting to summarize, rewrite, classify, extract information or analyze an attached reference text, ask open questions, brainstorming or chat with the model in the ZA locale, context, and personal preferences and provide the responses - Uploading an image and writing a prompt for the image based on various use cases, then writing a response for that prompt as though I were the chatbot. - Generate prompt and response pairs for a safety project

- Rate two model generated responses separately across various dimensions. Provide preference ranking scores. Write a justification for the response preference ranking and explain the specifics and logic behind the decision - Evaluate two AI responses to prompts designed to help provide services to business users - Create data on how AI models should handle user interactions that deal with sensitive topics or harmful themes. - Write prompts requesting to summarize, rewrite, classify, extract information or analyze an attached reference text, ask open questions, brainstorming or chat with the model in the ZA locale, context, and personal preferences and provide the responses - Uploading an image and writing a prompt for the image based on various use cases, then writing a response for that prompt as though I were the chatbot. - Generate prompt and response pairs for a safety project

2024

Media Search Analyst

TelusVideoEvaluation Rating

- Judging and grading Apple TV+ queries and results - Assess videos/movies for popularity and relevancy - Evaluate Apple tv+ content, ensuring search result accuracy and relevance - Utilize my cultural knowledge and market expertise to validate query intent in the ZA locale

- Judging and grading Apple TV+ queries and results - Assess videos/movies for popularity and relevancy - Evaluate Apple tv+ content, ensuring search result accuracy and relevance - Utilize my cultural knowledge and market expertise to validate query intent in the ZA locale

2025 - 2025

Education

Work History