Calvin Hizkia

Key Skills

Software

Appen

Labelbox

Remotasks

Scale AI

Top Subject Matter

No subject matter listed

Top Data Types

Audio

Image

Text

Top Task Types

Fine-tuning

RLHF

Translation/Localization

Freelancer Overview

I specialize in refining large language models through data annotation and RLHF evaluation, having worked on projects for leading AI labs like Scale AI. My core focus is on improving model response quality, naturalness, and multilingual accuracy. My effectiveness stems from a strong blend of technical annotation skills and bilingual expertise (English-Indonesian), which allows me to ensure culturally nuanced and precise data labeling for AI training. This is further supported by a sharp analytical mindset honed in finance.

Labeling Experience

Voice Activity Detection Annotator

LabelboxAudioAudio Recording

This project involved annotating audio data to train and evaluate AI speech recognition models. The core task was to meticulously transcribe and timestamp all speech and significant non-speech events in human-AI dialogues. Key Responsibilities: Precisely labeled speaker turns (User and AI Assistant) with millisecond accuracy. Classified user tokens into Standard Speech, Interruptions (barge-ins), and Acknowledgements (backchanneling like "mm-hmm"). Annotated non-speech events such as <pause>, <stop>, and <noise> (e.g., coughs, laughter) according to strict linguistic guidelines. Ensured continuous, gap-free annotation across the audio timeline to create seamless data for model training. Skipped tasks where the AI responded in a non-English language to maintain data integrity for the target models. The resulting annotated data was critical for improving how AI assistants handle natural conversational nuances like filled pauses, overlaps, and turn-taking.

This project involved annotating audio data to train and evaluate AI speech recognition models. The core task was to meticulously transcribe and timestamp all speech and significant non-speech events in human-AI dialogues. Key Responsibilities: Precisely labeled speaker turns (User and AI Assistant) with millisecond accuracy. Classified user tokens into Standard Speech, Interruptions (barge-ins), and Acknowledgements (backchanneling like "mm-hmm"). Annotated non-speech events such as <pause>, <stop>, and <noise> (e.g., coughs, laughter) according to strict linguistic guidelines. Ensured continuous, gap-free annotation across the audio timeline to create seamless data for model training. Skipped tasks where the AI responded in a non-English language to maintain data integrity for the target models. The resulting annotated data was critical for improving how AI assistants handle natural conversational nuances like filled pauses, overlaps, and turn-taking.

2025

Linguistic Evaluator

Scale AITextFine Tuning

As a Linguistic Evaluator for Meta's Multilingual Static Comparison Project V2, I specialized in enhancing AI model performance for the Indonesian language. My core responsibilities involved conducting comparative analysis of AI responses to identify the most accurate and contextually appropriate reply. I meticulously tagged errors in language, conversation flow, and cultural alignment, and ranked responses based on strict quality guidelines. For substandard outputs, I authored high-quality rewrites that corrected errors, improved fluency, and ensured cultural and contextual relevance, directly contributing to the model's training and refinement.

As a Linguistic Evaluator for Meta's Multilingual Static Comparison Project V2, I specialized in enhancing AI model performance for the Indonesian language. My core responsibilities involved conducting comparative analysis of AI responses to identify the most accurate and contextually appropriate reply. I meticulously tagged errors in language, conversation flow, and cultural alignment, and ranked responses based on strict quality guidelines. For substandard outputs, I authored high-quality rewrites that corrected errors, improved fluency, and ensured cultural and contextual relevance, directly contributing to the model's training and refinement.

2025 - 2025

Education

Work History