📍 Brazil (UTC-3)  |  🧪 10+ years in Software Quality  |  🤖 Stateful AI Agents & Local LLMs  |  📊 Data Analysis & ML

💼 Open to Remote Global Opportunities: Senior QA Engineer / SDET / Data QA Specialist / AI Test Engineer. 🎓 Currently pursuing an MBA in Applied Artificial Intelligence (UNISINOS, 2026).

I am a Senior Quality Engineer / SDET with 10+ years of experience, evolving from advanced test automation (Playwright, Cypress, Selenium) to Python data pipelines, predictive models and AI agents that run locally.

My core question:

How can we use AI and Data Science to move from reactive testing to predictable QA, without compromising data governance?

My QA rigor, rooted in reproducibility, strict validation and data integrity, shapes how I approach Data Science: I don't just consume AI models, I audit them with explicit metrics.

Instead of isolated repositories, my work is organized as an interconnected AI & QA data ecosystem. Each project covers one technical pillar (AI Engineering, Data Science or DevOps) of a continuous delivery lifecycle.

Status legend: ✅ Available  ·  🚧 In development / R&D  ·  💡 Planned

flowchart TB
    subgraph L1 ["Shift-Left Generation"]
        Gen["Test Case Generation<br/>RAG and Local LLMs<br/>from Business User Stories"]
    end

    subgraph L2 ["Execution and Telemetry"]
        direction LR
        Suites["Test Suites<br/>Cypress, Playwright,<br/>Selenium, API"] -->|"executed in"| Infra["CI/CD Pipeline<br/>Jenkins, Docker,<br/>SonarQube"]
    end

    subgraph L3 ["Intelligent Operations and Analytics"]
        direction LR
        ACV["Automation Code Validator<br/>LangGraph and Ollama"] -->|"Validation data"| Dash["Dashboard<br/>Metrics Quality"]
        Dash -->|"Consolidated telemetry"| Pred["Software Error<br/>Prediction Model"]
    end

    L1 -->|"Generated test scripts"| L2
    L2 -->|"Failure logs, metrics and telemetry"| L3
    L3 -.->|"Predictive risk feedback"| L1

    style ACV fill:#1f2d3d,stroke:#38bdf8,stroke-width:2px
    style Dash fill:#111827,stroke:#10b981,stroke-width:1px,stroke-dasharray: 5 5
    style L1 fill:transparent,stroke:#374151,stroke-width:1px,stroke-dasharray: 3 3
    style L2 fill:transparent,stroke:#374151,stroke-width:1px,stroke-dasharray: 3 3
    style L3 fill:transparent,stroke:#374151,stroke-width:1px,stroke-dasharray: 3 3

• Ecosystem role: The genesis layer. Runs before pipeline execution, turning product owner and QA items into BDD scenarios and test method skeletons.
• Engineering: A Python web application that uses RAG and a local LLM (Ollama) to interpret requirements, generate Gherkin scenarios and produce automation code for Cypress, Playwright, Selenium and API tests, with PostgreSQL persistence planned.
• Feedback loop: Designed to take failure-risk signals from the Software Error Prediction phase to prioritize coverage.
• Professional counterpart: Performance Improvement in Test Case Generation, an enterprise solution built with LLMs and MCP. Its source code is proprietary and not public due to confidentiality agreements. This repository is an independent concept on a public stack.

• Ecosystem role: The execution and telemetry engine. Containerizes, runs and monitors test suites, collecting unified telemetry.
• DevOps orchestration: Docker Compose environment (Jenkins, parallel runners for Playwright, Cypress and API tests, and a target application) for consistent environments across runs.
• Data aggregation: SonarQube Quality Gates inside Jenkins enforce code governance. The pipeline aggregates logs, videos, screenshots and JUnit XML metrics, forming the core dataset for the AI modules.
• Test repositories: e2e-cypress-js · e2e-playwright-python · api-testing

• Ecosystem role: The failure-triage layer. Analyzes automation logs and recommends fixes without manual log reading.
• AI architecture: A stateful agentic workflow built with LangGraph in Python (Parser → Evaluator → Fixer), running open-source LLMs locally via Ollama. It classifies failures such as ambiguous or flaky selectors and infrastructure timeouts, and recommends non-destructive locator fixes, falling back to NO_SAFE_FIX when context is ambiguous.
• Deterministic output: Strict Pydantic schemas and heuristics validate and override raw LLM output.

• Ecosystem role: The continuous feedback loop. Analyzes historical results to anticipate failures.
• Machine learning: A predictive pipeline trained on CI/CD telemetry logs, estimating failure probability before deployment.
• Closing the loop: Feeds failure patterns back into test case generation to focus coverage on historically error-prone rules and paths.

• Ecosystem role: The unified insights hub, consolidating telemetry from generation, execution and prediction.
• Engineering focus: An interface that translates QA metrics, data drift and automation coverage into business-ready health indicators.

A standalone repository with three reproducible notebook projects: unsupervised cybersecurity analytics, a deep reinforcement learning benchmark, and biomedical image segmentation, each with structured metrics.

• The challenge: Processing noisy, unlabeled firewall logs to identify stealthy infrastructure attacks.
• Data engineering: Shifted the analysis from raw packets to a behavioral matrix aggregated per Source IP, using Python and SQL.
• Statistical cleansing: Low-variance filtering plus automated multicollinearity pruning (VIF / correlation threshold) to remove geometric bias from distance calculations.
• Multi-model evaluation: K-Means, Hierarchical Clustering (Ward, with cophenetic validation) and DBSCAN (isolating malicious IPs as noise).
• Visualization: Interactive 2D PCA and t-SNE projections in Plotly, in a SOC-dashboard style.

• Architecture: DQN, PPO and A2C compared with Stable-Baselines3 and Gymnasium over 300k timesteps and independent seeds (discrete-action environment).
• Data pipeline: In-memory training stream with adaptive reward normalization (VecNormalize).
• Model auditing: Shannon Entropy tracking to detect policy degradation and policy collapse.

• Architecture: U-Net with multi-scale skip connections for joint-space boundary tracking in knee radiographs.
• Hardware optimization: VRAMKneeDataset GPU caching layer and target-aware augmentations with torchvision.transforms.v2.
• Metrics: 0.9569 mIoU and 0.9772 Dice on the held-out test set.

• Feature cleanliness over distance: In clustering pipelines, pruning highly correlated variables matters; redundant signals warp distance-based algorithms like K-Means.
• Model evaluation beyond accuracy: Auxiliary metrics such as Shannon action entropy reveal policy traps that reward tracking alone misses.
• Static governance prevents debt: Running code scanners (SonarQube) inside CI/CD keeps brittle test code and fragile locators out of the main branch.
• Retry-ability over fixed waits: Replacing arbitrary pauses (like cy.wait) with asynchronous state polling substantially stabilizes multi-container test infrastructure.

• MBA in Applied Artificial Intelligence, UNISINOS (in progress, 2026)
• Postgraduate Degree in Project Management, UNISINOS (2022)
• Bachelor's Degree in Information Systems, UNISINOS (2010)

• 💼 LinkedIn: linkedin.com/in/acorvello
• 📧 Email: alexsander.corvello@gmail.com
• 🐙 GitHub: @acorvello

Core competencies: Test Automation (Playwright, Cypress, Selenium) · CI/CD (Jenkins, Docker, SonarQube) · AI Agents (LangGraph, Ollama, MCP) · Machine Learning (scikit-learn, PyTorch) · Data Quality