Miguel Ángel Ávila González

Key Skills

Software

Appen

Scale AI

Toloka

Top Subject Matter

No subject matter listed

Top Data Types

Document

Image

Text

Top Task Types

Classification

Object Detection

RLHF

Text Generation

Translation Localization

Freelancer Overview

I have a solid track record as a bilingual AI Trainer and Data Entry Analyst, maintaining an accuracy rate of over 90% while processing large datasets. During my time at Toloka, I specialized in labeling and categorizing information to optimize machine learning models, in addition to implementing quality checks to maintain data integrity. This technical foundation is further supported by my background in Informatics, which allows me to understand the underlying mechanics of the models I train. My profile is distinguished by my advanced experience at Outlier AI, where I have focused on Large Language Model (LLM) projects. I have specialized in prompt engineering, crafting tailored instructions and performing comparative evaluations of generated responses to refine model quality. Furthermore, I possess critical skills in chat history analysis and fact-checking, ensuring that the AI provides responses that are not only coherent but also truthful. My ability to review specialized information provided by users ensures that models maintain a high standard of utility and safety.

Labeling Experience

Personalization Quality

Scale AITextClassificationQuestion Answering

The scope of the project was evaluating Meta AI so it could produce personalized answers using personal information of the user naturally. I read the whole conversation history between a user and the model to have the context needed, and evaluated if some specific information from a set was used and how efficiently was it used. The instructions and work process were simple. The project was a medium one, as it had hundred of contributors and lasted approximately 1 month.

The scope of the project was evaluating Meta AI so it could produce personalized answers using personal information of the user naturally. I read the whole conversation history between a user and the model to have the context needed, and evaluated if some specific information from a set was used and how efficiently was it used. The instructions and work process were simple. The project was a medium one, as it had hundred of contributors and lasted approximately 1 month.

2025 - 2025

Kestrel Nexus

Scale AITextClassificationEvaluation Rating

The scope of the project was evaluating Meta AI so it could produce factual answers grounded in sources. I read the whole conversation history between a user and the model to have the context needed to understand the intent of the current prompt and then evaluate the relevance and factualness of the answer. There were specific and changing instructions according to the needs of the client with complex rubrics to assess the factualness of each type of claim, including cases where the provided source was no linger available on the internet. The justifications for the factualness score were known as 'honesty proofs', and were detailed and covered each claim on the response. The project was a big one, as it had thousands of contributors and lasted approximately 5 months. We had to stick to strict rubrics, were evaluated and given feedback and had to participate in office hours.

The scope of the project was evaluating Meta AI so it could produce factual answers grounded in sources. I read the whole conversation history between a user and the model to have the context needed to understand the intent of the current prompt and then evaluate the relevance and factualness of the answer. There were specific and changing instructions according to the needs of the client with complex rubrics to assess the factualness of each type of claim, including cases where the provided source was no linger available on the internet. The justifications for the factualness score were known as 'honesty proofs', and were detailed and covered each claim on the response. The project was a big one, as it had thousands of contributors and lasted approximately 5 months. We had to stick to strict rubrics, were evaluated and given feedback and had to participate in office hours.

2025 - 2025

Cypher RLHF

Scale AITextText GenerationText Summarization

The scope of the project was training the AI model to produce high-quality answers in areas such as travelling, history, question answering, brainstorming, creative writing, between others. I reviewed two answers and looked for hallucinations, unfactual claims, risky content, misspellings and grammar errors, providing proofs and justifications for each error. Then I chose which answer was better and explained why. The project was a big one, as it had thousands of contributors and lasted approximately 4 months. We had to stick to strict rubrics and were evaluated and given feedback.

The scope of the project was training the AI model to produce high-quality answers in areas such as travelling, history, question answering, brainstorming, creative writing, between others. I reviewed two answers and looked for hallucinations, unfactual claims, risky content, misspellings and grammar errors, providing proofs and justifications for each error. Then I chose which answer was better and explained why. The project was a big one, as it had thousands of contributors and lasted approximately 4 months. We had to stick to strict rubrics and were evaluated and given feedback.

2024 - 2025

Map recognition

TolokaImageQuestion Answering

The project was about answering the directions a vehicle could drive from a specific point.

The project was about answering the directions a vehicle could drive from a specific point.

2023 - 2023

Signal recognition

TolokaImageBounding BoxClassification

I had to identify street signals and vehicles in a set of images.

I had to identify street signals and vehicles in a set of images.

2023 - 2023

Show more

Education

Work History