PROJECTS
View more →
Ctrl + LAB
A full-stack lab management system with a RAG-powered AI assistant.
FOLIVIX
An end-to-end system with a ViT model and Android app for maize leaf disease classification.
Gender Face Classifier
A classical ML project comparing Perceptron and SVM models for gender classification from facial images.
Satellite Image Classifier
A custom CNN for classifying high-resolution satellite imagery into 10 land cover types.
Face Detection
An adaptive face detection system improved with Reinforcement Learning from Human Feedback (RLHF).
Voice Classifier
A GUI-based system for training and testing KNN and SVM models on audio classification tasks.
S K I L L S
Technical
Soft
P R O J E C T S
Ctrl + LAB: Intelligent Lab Management System
A full-stack, responsive web application built to streamline academic lab management. It provides teachers with tools for scheduling, booking, and managing materials, while admins get full system oversight. Its core feature is 'Controly', a RAG-powered AI assistant using LangChain and a local LLM (Ollama) to answer contextual user queries. The entire system is containerized with Docker for seamless setup and deployment.
GitHubSkills
Results
FOLIVIX: Maize Leaf Disease Classifier
An end-to-end solution for maize leaf disease classification, featuring a Vision Transformer (ViT) model with 95% test accuracy. The system includes a Python backend with a Flask API for model serving and a native Android client, FOLIVIX, built with Kotlin and Jetpack Compose for real-time, on-the-field diagnosis. The app provides users with instant predictions, a filterable analysis history, and an in-app educational guide.
GitHubSkills
Results
Face Detection with RLHF
An advanced face detection system that integrates Reinforcement Learning from Human Feedback (RLHF) to continuously improve its performance. Built on a MobileNetV2 architecture, this project goes beyond static models by creating an adaptive learning loop. A custom GUI allows users to provide real-time feedback, which is used to retrain the model on challenging cases, achieving a 57% improvement in bounding box precision and a 64% reduction in overall loss.
GitHubSkills
Results
People Detection
This project showcases a classical machine learning approach to people detection, utilizing an Artificial Neural Network (ANN) trained on a rich set of custom-extracted features. Instead of relying on deep learning for feature extraction, this method demonstrates a deep understanding of feature engineering—using HOG, GLCM, and color metrics to classify image patches. The system is integrated into a custom GUI for real-time analysis of urban environments.
GitHubSkills
Results
Face Gender Classifier (Perceptron vs. SVM)
This project implements gender classification from facial images using two distinct classical machine learning approaches: a Neural Network Perceptron and a Support Vector Machine (SVM). Both models achieve over 90% test accuracy through an extensive feature engineering pipeline that combines color, texture (GLCM), shape (HOG), and geometric analysis. The project includes a user-friendly GUI built with CustomTkinter that enables real-time testing and side-by-side model comparison.
GitHubSkills
Results
Satellite Image Classifier
This project leverages a custom Convolutional Neural Network (CNN), 'SpectrumNet', to classify high-resolution satellite imagery from the EuroSAT dataset into 10 distinct land cover categories. The methodology includes robust data preprocessing, data augmentation, and a weighted loss function to handle class imbalance, achieving a 96% classification accuracy. The project is delivered with a user-friendly GUI built with Tkinter for real-time classification.
GitHubSkills
Results
Is this your voice? - Voice Classifier
This application provides a complete system for voice classification using KNN and SVM models. It features an intuitive GUI built with customtkinter that allows users to train models with custom hyperparameters, evaluate performance with confusion matrices and classification reports, and perform real-time classification on recorded or uploaded audio files. The system processes audio by extracting a rich set of acoustic descriptors like MFCCs, Spectral Centroid, and Zero-Crossing Rate.
GitHubSkills
Results
Animal Classifier (ResNet50)
An image classification system built using transfer learning with the ResNet50 model. This project fine-tunes the network to distinguish between cats, dogs, and snakes, achieving an impressive 98.67% accuracy. It features a custom Tkinter GUI for real-time, interactive predictions and employs robust data augmentation to ensure the model generalizes well to new, unseen images.
GitHubSkills
Results
E D U C A T I O N
University
Curriculum Map
Semester 1
Semester 2
Semester 3
Semester 4
Semester 5
Semester 6
Semester 7
Semester 8
Extracurricular Activities
Languages
Spanish
NativeEnglish
B2Chinese
HSK 1Courses
ABOUT ME
Professional Experience
Data Scientist
Outlier AI | 2025 - 2026 | LLM Evaluation for Data Science Workflows
Contributed to the improvement of LLMs in data science workflows. I designed complex tasks involving large-scale datasets that required precise coding, statistical operations, and high-fidelity visualizations. After prompting LLMs to generate solutions, I evaluated their outputs using structured rubrics, identifying logical gaps, implementation errors, and deviations from constraints. When models failed, I produced an ideal, step-by-step reasoning and implementation path the model should follow to get the correct response (numerical outputs and visualizations), ensuring strict adherence to requirements.
PythonpandasnumpymatplotlibseabornscipysklearnLLM promptingrubric-based evaluationdata visualizationGoogle ColabJupyter NotebooksCSV handlingJSON handlingXLSX handlingData Scientist
Outlier AI | 2025 | Spreadsheet & Structured Data Intelligence
Contributed to the evaluation and correction of LLMs designed to generate spreadsheet-based solutions from large, structured datasets. My role involved stress-testing models against detailed data requirements, auditing formula accuracy, logical consistency, and compliance with formatting constraints. When models failed, I reconstructed the correct analytical pathway, delivered the accurate solution, and documented failure modes. Using granular evaluation rubrics, I ensured every output met strict standards of numerical precision and structural correctness, reinforcing model reliability in enterprise-grade data workflows.
PythonpandasnumpyopenpyxlLLM promptingrubric-based evaluationdata analysisCSV handlingJSON handlingXLSX handlingSoftware Engineer
Outlier AI | 2025 | Autonomous Code Repair Systems
Contributed to the development of autonomous code-repair agents by generating high-quality repair trajectories across real-world GitHub repositories. I manually simulated the agent’s end-to-end reasoning process: analyzing issue reports, inspecting repository architecture, reproducing failures, identifying root causes, designing compliant patches, and validating fixes through rigorous unit and custom test construction. These structured trajectories served as supervised training data to teach the agent how to follow strict software engineering best practices, including clean code principles, robust test coverage, and reproducible Linux-based execution workflows.
PythonGitHubpytestLinux command linecode reviewdebuggingtestingpatchingpandasnumpyother librariesSoftware Engineer
Outlier AI | 2025 | LLM Code Generation Refinement
Focused on improving LLM-generated code through iterative refinement cycles. I evaluated model outputs against strict instruction sets, performance constraints, efficiency requirements, and formatting standards. When deficiencies were detected, I manually refactored and optimized logic, corrected edge cases, and enhanced computational performance. This multi-turn feedback loop strengthened the model’s ability to produce clean, efficient, and instruction-compliant code, driving outputs toward production-level reliability.
Pythoncode reviewoptimizationpandasnumpyseabornmatplotlibscipysklearnother librariesAI Engineer
Outlier AI | 2026 | Agent Evaluation & MCP Tooling
Evaluated and enhanced autonomous AI agents operating within tool-integrated environments (MCP systems), including file systems, Git, databases, mapping tools, and code execution frameworks. I designed adversarial and high-difficulty scenarios to identify reasoning breakdowns, trajectory failures, and tool misuse. For each failure, I engineered the ideal execution pathway, corrected outputs, and developed structured rubrics to systematically improve performance. This work significantly increased agent robustness, task completion reliability, and multi-tool reasoning coherence.
AI AgentsPythonMCPLinuxJSON handlingAPIsdatabasescode execution environmentsAI Engineer
Outlier AI | 2025 | Autonomous Build & Debugging Agents
Contributed to the refinement of AI agents responsible for diagnosing failed software builds and broken repositories. The agents analyzed execution logs, dependency errors, and test failures to determine root causes and implement corrective patches. When the agent's reasoning diverged from best practices or when failed to solve the problem, I reconstructed the correct debugging trajectory, improved scripts, reinforced unit and custom test validation, and ensured successful rebuilds. This role required deep understanding of CI/CD workflows, repository architecture, and automated debugging methodologies.
AI AgentsNode.jsCLICI/CDDockerGitLinuxGitHubpythonJavascriptKotlinother languagestesting frameworksMathematician
Outlier AI | 2024 - 2025 | LLM Mathematical Reasoning Specialist
Improved mathematical reasoning capabilities of Large Language Models across diverse domains, including algebra, calculus, geometry, and advanced problem-solving. I designed complex natural prompts to stress-test logical rigor and symbolic correctness. When models produced incorrect or incomplete solutions, I reconstructed fully rigorous mathematical derivations that strictly followed formal principles and constraints. My contributions strengthened symbolic accuracy, step-by-step reasoning integrity, and compliance with mathematical standards.
LaTeXGeogebraSymbolabalgebracalculusgeometryarithmeticnumber theorycombinatoricsother mathematical domainsFull Stack Developer
Private Enterprise Clients | Confidential
Designed and delivered end-to-end web applications integrating full-stack development with embedded data science components. I architected scalable backend systems optimized for high-throughput data processing, engineered efficient database querying strategies, and built interactive, analytics-driven frontend interfaces. Additionally, I implemented containerized environments and CI/CD pipelines to ensure reproducibility, deployment stability, and infrastructure consistency. All projects were developed under strict confidentiality agreements.
PythonJavaScriptReactNode.jsSQLDockerdockerGitHubSQLAlchemypandasnumpyCSV handlingJSON handlingGit
My Mission
My goal extends beyond engineering; I aspire to be an architect of the future. I leverage my skills in Artificial Intelligence not just to solve complex problems, but to design systems that fundamentally enhance our quality of life. I believe in a future where AI acts as a force for the common good, and my mission is to actively contribute to building that better, more intelligent world.
My Approach
Long-Term Vision
I build solutions that not only solve today's problem but are designed to scale and evolve into the future.
In-Depth Analysis
Before implementing a solution, I conduct a detailed analysis of the problem, considering various scenarios to choose the optimal path based on needs, resources, and time constraints.
World Impact
My motivation is to use AI to create a positive and meaningful impact on society and industry.
Beyond the Code
- Self-taught creator, editing engaging videos for my social media @alvarovasquez.ai.
- Passionate about music, playing instruments and exploring sound.
- Committed to physical discipline and strength training at the gym.
- Strategic thinker, honing my skills through competitive chess.
- Exploring the future of the decentralized web with Web 3.0 and Blockchain technology.
Academic Foundation
My education at the National Polytechnic Institute has provided me with a fundamental understanding of Artificial Intelligence. I not only master the mathematical theory and algorithms that drive the models, but also their practical application across diverse fields, their integration into complex systems, and the crucial ethical considerations that ensure responsible, human-centric technological development.
SOCIAL MEDIA
Alvaro Vasquez
@alvarovasquez.ai
My Objective
We will live in a world transformed by AI, and knowledge is our greatest tool. Often, the theory and formulas are a barrier. My objective is to tear it down using intuitive animations, allowing anyone to understand the fundamentals of AI from scratch to adapt and thrive in this new era.
Featured Content