01
Training & preference
Compare model responses, rank preferences, and write evidence-based rationales. Create golden responses and rewrite outputs for RLHF and supervised fine-tuning.
SENIOR AI TRAINER & LLM
EVALUATORCOLOMBIA · WORKING GLOBALLY
Human judgment.
Better
AI
I’m Juan Zapata. I evaluate how AI thinks, follows instructions, and writes code — turning human expertise into better training data.
Senior AI Trainer. Peer Reviewer. Software developer.
From
production code to the details that make a model better.
01 / MODEL OUTPUT
Human
judgmentREVIEW · TEST · REFINE
judgmentREVIEW · TEST · REFINE
02 / TRAINING SIGNAL
Clear feedback. Better data.
HUMAN-IN-THE-LOOP, BY DESIGN.
AI POST-TRAINING × SOFTWARE ENGINEERINGSCROLL TO EXPLORE
325+
Production tasks peer-reviewed
INVISIBLE TECHNOLOGIES
500+
AI training & evaluation tasks
SCALE AI / OUTLIER
120+
Frontend training projects overseen
TECHNICAL SQUAD LEAD
5+
Years across AI & engineering
BUILDING SINCE 202101 / EXPERTISE
Good models need
great human feedback.
I bring engineering depth to AI post-training: evaluating outputs, writing reference responses, and keeping human judgment consistent.
02
AI coding evaluation
Run generated code, reproduce bugs, and write tests. Assess correctness, instruction following, security, and maintainability across frontend, backend, and refactoring tasks.
03
Evaluation quality
Author verifiable rubrics, review contributor work, and resolve ambiguous cases. Calibrate reviewers and surface model failure modes through adversarial evaluation.
THE DEVELOPER’S PERSPECTIVE
Looks right.
But does it
actually work?
Code evaluation goes beyond a convincing answer. I check the requirements, execute the solution, test edge cases, and explain the result.
01 / INSPECT02 / EXECUTE03 / EXPLAIN
Try this small example. An original illustration of my review approach, using synthetic code.
evaluation / unique-values.jsJS
TASK
Return unique values, preserving their original types.
PAIRWISE REVIEW
// Candidate implementation
function unique(values) {
return [...new Set(
values.map(String)
)];
}// Candidate implementation
function unique(values) {
return [...new Set(values)];
}
// No type conversion
INPUT: [1, "1", 1, null]
Ready for review
Run the checks to inspect correctness and type preservation.
LOCAL, DETERMINISTIC CHECKSNO MODEL CONNECTED
02 / EXPERIENCE
From building software
to improving AI.
1.5+ years in AI post-training, grounded in hands-on production engineering.
Invisible Technologies
Freelance AI Trainer & Peer Reviewer RLHF / LLM Evaluation
Evaluate model responses through pairwise comparison, preference ranking, and multi-dimensional scoring. Advanced to Peer Reviewer, auditing 325+ production tasks for accuracy, instruction following, label quality, and rubric compliance.
- Author and review golden responses across correctness, relevance, completeness, style, and safety.
- Approve, correct, or return submissions with structured feedback; resolve scoring and guideline interpretation errors.
- Calibrate reviewer judgment, adjudicate ambiguous cases, and surface failure modes through selected red-teaming tasks.
Scale AI / Outlier
Technical Squad Lead & Senior Generative AI Engineer
Produced RLHF and SFT data across coding and general AI workflows. Led 5+ engineers, oversaw 120+ frontend training projects, and contributed to 500+ training and evaluation tasks.
- Built 60+ functional JavaScript and TypeScript frontend applications as training and evaluation artifacts.
- Executed code, reproduced bugs, authored tests, and compared solutions with evidence-based preference rationales.
- Designed evaluation rubrics and adversarial prompts; standardized reviews, improving team efficiency by approximately 30%.
Millennium Center
Full-Stack Developer
Built enterprise software, REST APIs, and microservices for shifts, payroll, and employee workflows with React/Next.js, Node.js, TypeScript, Docker, and MySQL. Process automation reduced operational costs by 35% and improved response times by 42%.
Energía Solar Helios
Full-Stack Developer & Data Analyst
Developed production platforms, APIs, database integrations, and automation using JavaScript, TypeScript, React/Next.js, Node.js, PHP, MySQL, and AWS S3. Improved application-processing efficiency by 45% and reduced operational errors by 60%.
TJDATA Group SAS
Full-Stack Developer & IT Support
Developed an e-commerce platform with React/Next.js and PHP, configured CRM systems, and improved online sales conversion by 18%.
03 / SELECTED ENGINEERING WORK
Built it. Debugged it.
Now I know what to look for.
Production software from my engineering portfolio. The practical foundation behind how I review AI-generated code.
Shift rotations, payroll, and employee workflows in one enterprise management system.
Configurable WhatsApp conversation flows for sales, support, and database-connected responses.
City lighting maintenance, work orders, location lookup, and material inventory tracking.
Credit and recruitment workflows connected to contract management, payroll, and ERP/CRM.
04 / THE FOUNDATION
Technical depth.
Human perspective.
TOOLS I WORK WITH
JavaScriptTypeScriptPythonReact / Next.jsNode.js / ExpressREST APIsMySQLPHPGitDocker
EDUCATION & TRAINING
Information Systems Analysis
and Development
Associate Degree · SENA · 2022–2024
Oracle for Java MinTIC · 2022
Mobile Device Programming SENA · 2021
LANGUAGES & COLLABORATION
English Professional working proficiency
Spanish Native
Experienced in asynchronous,
English-language production teams.
English-language production teams.
05 / LET’S CONNECT
Better AI starts
with a good
conversation.
Looking for an AI trainer, coding evaluator, or peer reviewer with an engineering background? Let’s talk.
juanpazapata2013@gmail.com