Sakhina Crosslin Premkumar

Key Skills

Software

Mercor

Scale AI

Top Subject Matter

No subject matter listed

Top Data Types

Image

Text

Document

Video

Top Label Types

Classification

Question Answering

Text Generation

Emotion Recognition

Object Detection

Translation Localization

Evaluation Rating

Transcription

Text Summarization

Fine Tuning

Prompt Response Writing SFT

Audio Recording

Freelancer Overview

I am an experienced AI training data specialist with a strong focus on image-based Q&A review, data labeling, and answer correction. My work involves auditing and verifying labels and answers against detailed evidence in screenshots, documents, and charts, ensuring high accuracy and consistency as per project guidelines and rubrics. I am skilled in identifying defects, rewriting answers to be clear and standalone, and maintaining rigorous calibration standards. My background also includes prompt evaluation for LLMs, image/video generation prompt refinement, and content editing, all supported by C1-level written English and a keen eye for detail. I am proficient with tools like Parimango, Label Studio, Excel, and collaborative platforms such as MS Teams and Confluence, and I am comfortable meeting efficiency targets in remote, quality-driven environments.

Labeling Experience

AI Prompt Engineering

MercorVideoFine TuningPrompt Response Writing SFT

Designed and refined multimodal prompts for image/video generation (Gemini) and created language prompts to improve instruction-following, tone control, and natural phrasing across outputs. Iterated prompts based on observed failure modes, tested edge cases, and documented prompt patterns that improved consistency and quality. Evaluated outputs for constraint adherence (style, composition, entities, and safety boundaries) and language quality (clarity, grammar, fluency). Produced clean prompt rewrites and concise QA notes to support faster iteration and repeatable results.

Designed and refined multimodal prompts for image/video generation (Gemini) and created language prompts to improve instruction-following, tone control, and natural phrasing across outputs. Iterated prompts based on observed failure modes, tested edge cases, and documented prompt patterns that improved consistency and quality. Evaluated outputs for constraint adherence (style, composition, entities, and safety boundaries) and language quality (clarity, grammar, fluency). Produced clean prompt rewrites and concise QA notes to support faster iteration and repeatable results.

2025 - 2025

Math & Reasoning Evaluation

Scale AIDocumentText SummarizationTranslation Localization

Reviewed math and reasoning Q&A items to verify final answer correctness and step-by-step validity against the prompt requirements. Evaluated responses using rubric-based criteria, checking for arithmetic accuracy, logical consistency, constraint adherence, and clarity. Tagged common issues such as incorrect computations, missing steps, invalid assumptions, and formatting problems. When required, rewrote solutions into clear, standalone corrected answers with consistent notation and strong grammar, ensuring the output is objectively verifiable and instruction-aligned.

Reviewed math and reasoning Q&A items to verify final answer correctness and step-by-step validity against the prompt requirements. Evaluated responses using rubric-based criteria, checking for arithmetic accuracy, logical consistency, constraint adherence, and clarity. Tagged common issues such as incorrect computations, missing steps, invalid assumptions, and formatting problems. When required, rewrote solutions into clear, standalone corrected answers with consistent notation and strong grammar, ensuring the output is objectively verifiable and instruction-aligned.

2025 - 2025

Multilingual LLM Response Evaluation

Scale AITextTranslation LocalizationEvaluation Rating

Performed multilingual LLM response evaluation using structured guidelines and rubric-based scoring. Compared paired responses (A/B) and assessed grammar, fluency, natural phrasing, conversation fit, cultural appropriateness, and instruction adherence. Produced clear preference justifications with audit-ready rationale and applied defect tagging for common failure modes (language errors, missing constraints, incorrect tone, unclear answers, formatting issues). When required, rewrote responses into clean, user-ready outputs with English, ensuring clarity, correctness, and consistency with the task instructions. Maintained calibration alignment and quality consistency across batches.

Performed multilingual LLM response evaluation using structured guidelines and rubric-based scoring. Compared paired responses (A/B) and assessed grammar, fluency, natural phrasing, conversation fit, cultural appropriateness, and instruction adherence. Produced clear preference justifications with audit-ready rationale and applied defect tagging for common failure modes (language errors, missing constraints, incorrect tone, unclear answers, formatting issues). When required, rewrote responses into clean, user-ready outputs with English, ensuring clarity, correctness, and consistency with the task instructions. Maintained calibration alignment and quality consistency across batches.

2025 - 2025

Image-Understanding

MercorImageClassificationQuestion Answering

Delivered high-precision, schema-driven annotations on diverse images to improve multimodal model understanding and retrieval. Produced consistent metadata across multiple required fields—Named Entities, Unnamed Entities, Keywords, Activities, Emotional Tone, Editorial Intent, Topic/Genre, OCR text transcription, long Summary (150–300 words), and Short Summary (50–100 words)—while enforcing strict quality controls to prevent hallucinations, unsupported assumptions, and formatting drift. Work included rigorous policy screening and rare-case rejection handling (e.g., policy violations, non-rendering content), and all outputs were written in American English with tight grammar and spelling consistency.

Delivered high-precision, schema-driven annotations on diverse images to improve multimodal model understanding and retrieval. Produced consistent metadata across multiple required fields—Named Entities, Unnamed Entities, Keywords, Activities, Emotional Tone, Editorial Intent, Topic/Genre, OCR text transcription, long Summary (150–300 words), and Short Summary (50–100 words)—while enforcing strict quality controls to prevent hallucinations, unsupported assumptions, and formatting drift. Work included rigorous policy screening and rare-case rejection handling (e.g., policy violations, non-rendering content), and all outputs were written in American English with tight grammar and spelling consistency.

2025 - 2025

Education

Work History