Andrei Dobos

Key Skills

Software

Appen

Scale AI

Top Subject Matter

No subject matter listed

Top Data Types

Audio

Text

Top Label Types

Classification

RLHF

Prompt Response Writing SFT

Freelancer Overview

I am an experienced Knowledge Graph Specialist and academic editor with a strong background in linguistic analysis, data verification, and content curation. My recent work on a large-scale Knowledge Graph project involved semantic analysis of search queries, entity disambiguation, and OSINT-based data verification for a major search engine, where I ensured the accuracy and cultural relevance of AI training data in both English and Romanian. With over a decade of experience in managing complex editorial projects, translating technical content, and refining metadata, I bring a meticulous approach to data labeling and annotation, particularly in natural language processing and information extraction domains. My expertise in ontology, source validation, and multilingual data localization allows me to deliver high-quality, contextually precise training datasets for AI and machine learning applications.

Labeling Experience

Appen – Jigglypuff

AppenAudioClassification

Performed large-scale text annotation and quality evaluation within Appen's enterprise pipeline for a major AI client. Tasks included applying detailed annotation guidelines to classify and evaluate text data, maintaining high inter-annotator agreement, and ensuring consistency across thousands of labeled examples. Applied linguistic expertise (PhD in Philology, native Romanian, C2 English) to handle nuanced labeling decisions requiring semantic precision and cultural context awareness.

Performed large-scale text annotation and quality evaluation within Appen's enterprise pipeline for a major AI client. Tasks included applying detailed annotation guidelines to classify and evaluate text data, maintaining high inter-annotator agreement, and ensuring consistency across thousands of labeled examples. Applied linguistic expertise (PhD in Philology, native Romanian, C2 English) to handle nuanced labeling decisions requiring semantic precision and cultural context awareness.

2025

Outlier – Gemini Safety (Moldovan Elections)

Scale AITextClassification

Trained Google's Gemini model to identify and flag malicious propaganda in the context of Moldovan electoral disinformation. Evaluated model outputs for political bias, factual manipulation, and culturally specific misinformation patterns in Romanian-language content. Work involved adversarial testing and red-teaming requiring deep understanding of Eastern European political dynamics, Romanian/Moldovan linguistic nuance, and propaganda techniques. Applied safety-critical annotation standards with zero tolerance for false negatives on harmful content.

Trained Google's Gemini model to identify and flag malicious propaganda in the context of Moldovan electoral disinformation. Evaluated model outputs for political bias, factual manipulation, and culturally specific misinformation patterns in Romanian-language content. Work involved adversarial testing and red-teaming requiring deep understanding of Eastern European political dynamics, Romanian/Moldovan linguistic nuance, and propaganda techniques. Applied safety-critical annotation standards with zero tolerance for false negatives on harmful content.

2025 - 2025

Outlier – Barbeque_doe

Scale AITextRLHFPrompt Response Writing SFT

Conducted reinforcement learning from human feedback (RLHF) tasks for large language model training. Work included pairwise response ranking, preference labeling, and qualitative evaluation of model-generated text for factual accuracy, coherence, logical reasoning, and stylistic quality. Leveraged domain expertise in humanities and linguistics to evaluate complex outputs across multiple quality dimensions. Maintained high accuracy scores against gold-standard benchmarks across 500+ hours of evaluation work.

Conducted reinforcement learning from human feedback (RLHF) tasks for large language model training. Work included pairwise response ranking, preference labeling, and qualitative evaluation of model-generated text for factual accuracy, coherence, logical reasoning, and stylistic quality. Leveraged domain expertise in humanities and linguistics to evaluate complex outputs across multiple quality dimensions. Maintained high accuracy scores against gold-standard benchmarks across 500+ hours of evaluation work.

2024 - 2025

Education

Work History