Joan Cabanas

Key Skills

Software

Appen

Data Annotation Tech

Mercor

SuperAnnotate

Top Subject Matter

No subject matter listed

Top Data Types

Audio

Image

Text

Top Task Types

Evaluation Rating

Fine Tuning

Prompt Response Writing SFT

RLHF

Translation Localization

Freelancer Overview

I’m an experienced AI model evaluator and prompt engineer with a strong background in LLM output analysis, instruction tuning, and human-in-the-loop model training. I specialize in crafting high-quality reference responses, building guidelines for model alignment, and creating evaluation prompts that help AI systems reason more safely and effectively. With 15+ years of experience as a technical sales engineer prior to entering AI, I bring a unique ability to translate user intent, domain knowledge, and system requirements into precise, reliable model behavior.

Labeling Experience

Expert Prompt Solver

MercorTextQuestion Answering

Manually solved complex prompts requiring deep reasoning, research and information synthesis. I then ran the same prompts through advanced reasoning models such as GPT.5 and Gemini 2.5 Pro and compared their responses against my own solutions to evaluate reasoning accuracy and depth of understanding.

Manually solved complex prompts requiring deep reasoning, research and information synthesis. I then ran the same prompts through advanced reasoning models such as GPT.5 and Gemini 2.5 Pro and compared their responses against my own solutions to evaluate reasoning accuracy and depth of understanding.

2025

Rubrics/Guidelines

Data Annotation TechTextFine Tuning

Designed detailed instruction guidelines for AI models to interpret both direct and indirect instructions within a given prompt, ensuring accurate understanding of its objective and subjective dimensions. The goal was to curate a set of instructions capable of generating the most precise and contextually appropriate response.

Designed detailed instruction guidelines for AI models to interpret both direct and indirect instructions within a given prompt, ensuring accurate understanding of its objective and subjective dimensions. The goal was to curate a set of instructions capable of generating the most precise and contextually appropriate response.

2025

Prompt Engineering

Data Annotation TechTextQuestion AnsweringText Generation

Designed a diverse set of safe, non-adversarial prompts across multiple domains to evaluate model behavior under various conditions. For each prompt, the model produced two responses that I evaluated along multiple qualitative axis. I the provided detailed written analysis highlighting areas where the model excelled and where its performance was either poor or could be improved.

Designed a diverse set of safe, non-adversarial prompts across multiple domains to evaluate model behavior under various conditions. For each prompt, the model produced two responses that I evaluated along multiple qualitative axis. I the provided detailed written analysis highlighting areas where the model excelled and where its performance was either poor or could be improved.

2024

RLHF

Data Annotation TechTextRLHF

Evaluation two different model responses on different axis such as "Verbosity", "Instruction following", "Safety", "Localization", "Truthfulness" and "Writing quality". Providing a detailed explanation on where the model failed and where the model excelled.

Evaluation two different model responses on different axis such as "Verbosity", "Instruction following", "Safety", "Localization", "Truthfulness" and "Writing quality". Providing a detailed explanation on where the model failed and where the model excelled.

2024

Audio prompts creation

Data Annotation TechAudioAudio Recording

Designed and recorded a set of various prompts covering various topics. The prompts were recorded under different ambient conditions. The goal was to create a wide range of naturally sounding questions featuring real-life background noise, such as metro stations, a supermarket, and busy bars.

Designed and recorded a set of various prompts covering various topics. The prompts were recorded under different ambient conditions. The goal was to create a wide range of naturally sounding questions featuring real-life background noise, such as metro stations, a supermarket, and busy bars.

2025 - 2025

Education

Work History