Julian Narvaja

Key Skills

Software

Appen

CrowdSource

Data Annotation Tech

Labelbox

OneForma

Remotasks

Scale AI

Telus

Internal/Proprietary Tooling

Top Subject Matter

No subject matter listed

Top Data Types

Audio

Image

Text

Top Task Types

Action Recognition

Audio Recording

Data Collection

Emotion Recognition

Evaluation Rating

Freelancer Overview

I have hands-on experience in data annotation and AI training data creation, specializing in audio recognition, conversation labeling, and linguistic data preparation. My work includes labeling speech patterns, transcribing and tagging conversations, and creating structured datasets that improve the accuracy of speech-to-text and conversational AI systems. In addition, I have contributed to Large Language Model (LLM) evaluation and training, where I reviewed outputs, identified errors, and refined prompts to enhance overall model performance.

Labeling Experience

Xylophone Grassland/Conversation

Scale AIAudioClassificationEmotion Recognition

Listening to or reading conversations (human-human or human-AI). Labeling speaker turns, intent, and context (e.g., who is speaking, what they want, whether the answer is relevant). Tagging audio patterns like pauses, interruptions, or non-speech events when required. Evaluating model responses for clarity, accuracy, and alignment with the conversational flow. In some cases, generating or editing sample dialogues to train LLMs for more natural, human-like interaction. The end goal is to provide clean, structured training data that helps conversational models better understand dialogue dynamics, intent recognition, and natural flow in multiple languages.

Listening to or reading conversations (human-human or human-AI). Labeling speaker turns, intent, and context (e.g., who is speaking, what they want, whether the answer is relevant). Tagging audio patterns like pauses, interruptions, or non-speech events when required. Evaluating model responses for clarity, accuracy, and alignment with the conversational flow. In some cases, generating or editing sample dialogues to train LLMs for more natural, human-like interaction. The end goal is to provide clean, structured training data that helps conversational models better understand dialogue dynamics, intent recognition, and natural flow in multiple languages.

2025 - 2025

engie-reader-iqa

Scale AIDocumentBounding BoxSegmentation

Read passages and questions provided to the model. Check if the model’s answer is correct, complete, and faithful to the text. Flag errors like hallucinations, incomplete reasoning, or answers not grounded in the passage. Sometimes, write or refine gold-standard answers to improve the training data. The overall goal is to create high-quality evaluation data that helps train and benchmark question-answering and reading comprehension models. It’s less about raw labeling (like audio tagging) and more about judging model accuracy and data integrity, which requires attention to detail and critical reasoning.

Read passages and questions provided to the model. Check if the model’s answer is correct, complete, and faithful to the text. Flag errors like hallucinations, incomplete reasoning, or answers not grounded in the passage. Sometimes, write or refine gold-standard answers to improve the training data. The overall goal is to create high-quality evaluation data that helps train and benchmark question-answering and reading comprehension models. It’s less about raw labeling (like audio tagging) and more about judging model accuracy and data integrity, which requires attention to detail and critical reasoning.

2025 - 2025

Education

Work History