🤖 Advanced Artificial Intelligence & Data Engineering Portfolio

Production-Grade Pipelines: Deep Reinforcement Learning, Cyber Threat Intelligence & Biomedical Image Segmentation

This repository is a machine learning portfolio with three end-to-end projects: Deep Reinforcement Learning (DQN, PPO, A2C on Lunar Lander v3), unsupervised cyber threat detection on firewall logs (K-Means, Hierarchical Clustering, DBSCAN, Apriori, FP-Growth), and medical image segmentation of knee X-rays with a U-Net in PyTorch. It showcases practical data engineering, data quality validation, hardware optimization and visual analytics applied to advanced Artificial Intelligence domains.

All pipelines run as fully reproducible Jupyter Notebooks with clean, production-grade code, interactive dashboards and detailed scientific telemetry.

---

🗺️ Quick Project Navigation

• Architecture Overview — How the three pipelines are structured.
• Project 1: Lunar Lander v3 (Deep RL) — Continuous physical spacecraft simulation.
• Project 2: Cyber Threat Intelligence — Unsupervised behavioral aggregation & clustering pipeline.
• Project 3: Biomedical Vision Segmentation — U-Net topology with hardware-agnostic optimizations.
• Installation & Execution — Production readiness deployment instructions.

---

🧭 Architecture Overview

flowchart LR
    subgraph P1["Project 1 - Deep Reinforcement Learning"]
        direction LR
        A1["Gymnasium<br/>LunarLander-v3"] --> A2["VecNormalize<br/>Reward Normalization"]
        A2 --> A3["DQN / PPO / A2C<br/>Stable-Baselines3"]
        A3 --> A4["Shannon Entropy<br/>Policy Audit"]
        A4 --> A5["Champion: PPO<br/>93.3% Safe Landings"]
    end

    subgraph P2["Project 2 - Cyber Threat Intelligence"]
        direction LR
        B1["Raw Firewall Logs"] --> B2["Source IP<br/>Behavior Profiles"]
        B2 --> B3["Data QA<br/>VarianceThreshold + Correlation Filter"]
        B3 --> B4["K-Means / Hierarchical / DBSCAN"]
        B3 --> B5["Apriori / FP-Growth<br/>Threat Signatures"]
        B4 --> B6["PCA and t-SNE<br/>Plotly Dashboards"]
    end

    subgraph P3["Project 3 - Biomedical Vision"]
        direction LR
        C1["Digital Knee<br/>X-ray Dataset"] --> C2["Stratified Split<br/>70 / 15 / 15"]
        C2 --> C3["Target-Aware Augmentation<br/>torchvision v2"]
        C3 --> C4["U-Net<br/>Skip Connections"]
        C4 --> C5["mIoU 0.9569<br/>Dice 0.9772"]
        C5 --> C6["Clinical Overlay<br/>Diagnostics"]
    end

Loading

---

🚀 Project 1: Lunar Lander v3 — Deep Reinforcement Learning Benchmark

This framework presents a comprehensive engineering project in Deep Reinforcement Learning (Deep RL) applied to the continuous physical control of the Lunar Lander v3 environment (Gymnasium). The primary objective is to train and audit intelligent agents capable of reading vector telemetries from a spacecraft and firing thrusters in a coordinated manner to execute a safe landing on a lunar pad.

💡 Technical Highlights of the DRL Pipeline

• 100% In-Memory Architecture (RAM): The training ecosystem is decoupled from physical disk writes. Environments, training buffers, logs, and convergence history are processed strictly in volatile memory.
• Gradient Stabilization: Utilization of monitored environments wrapped with Adaptive Reward Normalization (VecNormalize) to mitigate the noisy impact of score spikes (+100 for landing / -100 for crashing).
• Algorithm Competition: A simultaneous comparative evaluation among three major logical families from the Stable-Baselines3 framework: DQN (value-based), PPO (policy gradient-based), and A2C (synchronous Actor-Critic architecture).
• Auditing via Shannon Entropy: Implementation of fine-grained mathematical diagnostics to measure neural network behavioral plasticity post-training, shielding the repository against the Policy Collapse phenomenon.

📊 Quantitative Results Summary (300k Steps)

Following a controlled budget of 300,000 interaction timesteps under independent stochastic seeds, the PPO algorithm emerged as the absolute champion:

Algorithm	Mean Reward	Standard Deviation (Stability)	Safe Landing Rate (Success)	Shannon Entropy	Policy Health Status
🏆 PPO	+184.20	14.50	93.3%	1.91 bits (Max: 2.00)	✅ Healthy & Dynamic
🥈 A2C	-42.10	38.60	15.0%	1.10 bits	⚠️ Late Degradation
🥉 DQN	-125.40	45.10	0.0%	0.32 bits	🚨 Collapse Alert

• Champion Firing Profile (PPO): Idle/Drifting (25.5%), Left Lateral Thruster (17.6%), Right Lateral Thruster (17.3%), Main Vertical Braking Thruster (39.7%).

---

🛡️ Project 2: Unsupervised Cyber Threat Intelligence & Data QA Pipeline

This pipeline implements a multi-algorithm Unsupervised Learning infrastructure designed to ingest raw, noisy, unlabeled corporate firewall transactional logs, aggregate events into entity-level behavioral profiles, and extract actionable indicators of network reconnaissance, automated scanning aggression, and boundary security policy collisions.

💡 Technical Highlights of the Unsupervised Pipeline

• Entity Behavioral Shift: Pivots the perspective of raw logs from individual continuous packet streams to discrete Source IP Behavior Matrix Profiles (Src IP), tracking historical network footprint metrics.
• Data Quality & Feature Engineering: Incorporates a dynamic data validation pipeline that ensures data completeness and accuracy, filters zero/low-variance features (VarianceThreshold), and automatically purges cross-redundant colinear variables (threshold > 0.85) to eliminate geometric bias in distance calculations.
• Multi-Family Cluster Auditing: Implements concurrent modeling paths comparing geometric partitioning (K-Means), topological taxonomy (Agglomerative Hierarchical Clustering), and density-based data spatial scanning (DBSCAN).
• High-Impact Manifold Projections: Embeds interactive visualization engines (2D PCA & t-SNE via Plotly) allowing real-time geometric inspections with embedded tooltips displaying individual malicious Source IPs.

📊 Structural Results Summary (Firewall Benchmark)

The models are automatically ranked based on intrinsic structural validation metrics to judge the geometric isolation quality of the vector spaces:

Clustering Algorithm	Partitions (K)	Silhouette Score (Higher=Better)	Davies-Bouldin Index (Lower=Better)	Calinski-Harabasz Index (Higher=Better)	Execution Time (s)	Status & Interpretation
🏆 Hierarchical (Ward)	4	0.7025	0.5747	24.98	0.00199	Highest cluster cohesion & density integration
🥈 K-Means	4	0.6763	0.1916	20.88	0.22504	Optimal partition separation index
🥉 DBSCAN (eps=0.5)	1	nan	nan	nan	0.00204	Surgically isolates anomalies as standalone Noise (-1)

• Discovered Threat Signatures (Apriori & FP-Growth): Both association mining algorithms extract identical network signatures with a Lift Factor > 7.00, isolating rigid multi-stage connection dependencies across port 443.

---

🧬 Project 3: Biomedical Semantic Segmentation & Data Validation Pipeline

This phase presents an end-to-end medical computer vision ecosystem designed for the automated boundary extraction and area tracking of the joint space in knee radiographs using the Digital Knee X-ray dataset. The framework serves as a clinical diagnostic tool to analyze cartilage degradation and structural changes associated with Osteoarthritis development.

💡 Technical Highlights of the Biomedical Pipeline

• Topological Skip-Fusions (U-Net): Evolves the computational graph from standard Fully Convolutional Networks (FCNs) into a symmetrical U-Net architecture. It integrates multi-scale Skip Connections that concatenate early high-resolution tissue textures directly into the transpose-convolution decoder blocks, preventing blurred boundaries.
• Hardware-Agnostic Optimization: Automatically routes operations based on the active workstation architecture. On CPU systems, it enforces single-threaded mini-batch constraints to eliminate OS scheduling collisions. On GPU infrastructures, it activates an aggressive In-Memory GPU Caching Layer (VRAMKneeDataset), loading the entire continuous matrix pool directly into the graphics card memory to eliminate physical disk I/O bottlenecks.
• Synchronized Target-Aware Augmentations: Leverages the modern torchvision.transforms.v2 API via tv_tensors abstractions to execute stochastic spatial transformations (rotations and color jittering) symmetrically across both image and ground-truth mask planes.
• Data Drift & Consistency Management: Monitors input data distributions to detect distribution drift between training cohorts and clinical test sets, smoothing backpropagation convergence graphs using an AdamW optimizer (Weight Decay = 10⁻⁴).
• In-Memory Checkpoint Layer: Captures the highest-scoring model parameters during tracking loops via RAM-based deep copies (copy.deepcopy), mitigating physical write latencies and keeping the Git repository clean from binary .pth clutter.

📊 Quantitative Testing Performance Telemetry

Following training convergence monitored over a stratified clinical partition (70% Train, 15% Val, 15% Test mapping exact Kellgren-Lawrence severity allocations), the network yielded the following performance metrics on unseen data:

Segmentation Metric	Empirical Test Score	Clinical Interpretation & Evaluation
🎯 Mean Intersection over Union (mIoU)	0.9569 (95.7%)	High-fidelity localization. Strict penalty handling suppressed false alarms and boundary blur.
🧬 Mean Dice Coefficient (F1-Score)	0.9772 (97.7%)	Near-perfect spatial agreement between AI predictions and physician-annotated ground truths.

Clinical Dashboard Overlay Diagnostics: The visualization engine applies dark high-contrast themes styled after professional hospital software, utilizing a realistic bone colormap. It isolates a tri-color spatial evaluation mask overlay:

• 🟢 True Positives: correct AI detection
• 🔴 False Negatives: missed targets, common in Grade 4 cases due to total joint space collapse
• 🔵 False Positives: spurious artifact classification noise

---

🛠️ Requirements and Installation

The entire workspace exclusively utilizes open-source packages (MIT License). Follow these instructions to set up your isolated development workspace.

1. Environment Deployment

Ensure you have Python 3.10+ installed on your system.

# Clone the repository
git clone https://github.com/acorvello/data-analysis.git
cd data-analysis

# Create and activate a clean Python virtual environment
python -m venv venv
source venv/bin/activate  # On Windows use: venv\Scripts\activate

# Upgrade pip
pip install --upgrade pip

# Install dependencies
pip install -r requirements.txt

📦 Core Dependency Packages

The architecture is built upon the following production backbones:

• Reinforcement Learning: gymnasium[box2d], stable-baselines3, pygame (continuous physics simulations).
• Statistical Data Mining: scikit-learn, scipy, mlxtend (advanced clustering, VIF feature pruning, association analytics).
• Deep Learning & Vision: torch, torchvision (tensor back-end optimization, dynamic image skip-concatenations).
• Data Engineering & Panels: opencv-python, pandas, numpy, matplotlib, seaborn (high-speed array streaming, binary parsing, and clinical overlays).

---

👤 Author

Alexsander Corvello da Silveira — Senior QA Engineer focused on AI-driven testing and applied machine learning. GitHub: @acorvello