Preyash Kharva

Key Skills

Software

CVAT

Data Annotation Tech

Labelbox

Mercor

Remotasks

Scale AI

Telus

Other

Top Subject Matter

No subject matter listed

Top Data Types

Document

Image

Video

Top Task Types

Data Collection

Fine Tuning

Prompt Response Writing SFT

Red Teaming

RLHF

Freelancer Overview

As an experienced AI Trainer and Data Labeling Specialist with over three years of hands-on experience, I have contributed to large-scale annotation, evaluation, and safety projects for leading AI companies including Outlier AI, Handshake AI, Abaka.ai, and NamanSoft Inc. My expertise spans LLM evaluation, multimodal annotation (text, image, audio, video), red teaming, and Reinforcement Learning from Human Feedback (RLHF). At Handshake AI, I led red-teaming and adversarial data tasks to identify vulnerabilities in frontier language models (GPT-class systems), ensuring compliance, ethical safety, and robustness. With Abaka.ai, I performed detailed quality reviews of global datasets, enhancing labeling consistency and accuracy by 18%. My work at Outlier AI focused on building structured datasets and prompt evaluation pipelines that improved model reasoning quality and factual accuracy. Currently at NamanSoft, I integrate labeled data and AI pipelines to power automation and agentic systems, directly improving AI system reliability and enterprise workflow efficiency.

Labeling Experience

Multimodal Data Labeling & Quality Assurance for AI Systems

Scale AIVideoClassificationObject Detection

At Abaka.ai, I contributed to large-scale multimodal labeling and evaluation tasks across text, image, video, and audio datasets used to train next-generation AI systems. Responsibilities included classifying, segmenting, and reviewing content for quality, bias, and accuracy, as well as conducting detailed QA checks to ensure dataset reliability and compliance. Worked collaboratively with annotation teams across multiple time zones, ensuring precise consistency across hundreds of data batches. Designed custom validation workflows that helped increase labeling precision by 18% and reduced annotation turnaround time.

At Abaka.ai, I contributed to large-scale multimodal labeling and evaluation tasks across text, image, video, and audio datasets used to train next-generation AI systems. Responsibilities included classifying, segmenting, and reviewing content for quality, bias, and accuracy, as well as conducting detailed QA checks to ensure dataset reliability and compliance. Worked collaboratively with annotation teams across multiple time zones, ensuring precise consistency across hundreds of data batches. Designed custom validation workflows that helped increase labeling precision by 18% and reduced annotation turnaround time.

2025

AI Red Teaming & Model Safety Evaluation (LLM Data Labeling Project)

CVATImageRLHFFine Tuning

As part of the Handshake AI MOVE Fellowship, I contributed to large-scale data labeling and evaluation tasks for GPT-class language models, focusing on red teaming, RLHF alignment, and multimodal prompt evaluation. I designed and executed adversarial prompts to uncover model weaknesses in data exfiltration, compliance bypass, and reasoning quality. The project involved manually labeling and reviewing thousands of model responses across multilingual contexts, annotating outputs for quality, bias, safety, and factual accuracy. Collaborated with applied AI researchers to design reproducible evaluation pipelines that improved model interpretability and robustness. (SOFTWARE USED -> HANDSHAK AI PARTNERED WITH OPEN AI, Platform used FEATHER)

As part of the Handshake AI MOVE Fellowship, I contributed to large-scale data labeling and evaluation tasks for GPT-class language models, focusing on red teaming, RLHF alignment, and multimodal prompt evaluation. I designed and executed adversarial prompts to uncover model weaknesses in data exfiltration, compliance bypass, and reasoning quality. The project involved manually labeling and reviewing thousands of model responses across multilingual contexts, annotating outputs for quality, bias, safety, and factual accuracy. Collaborated with applied AI researchers to design reproducible evaluation pipelines that improved model interpretability and robustness. (SOFTWARE USED -> HANDSHAK AI PARTNERED WITH OPEN AI, Platform used FEATHER)

2024

Education

Work History