Professional implementation of 15+ AI/ML projects spanning classical algorithms, supervised learning, neural networks, and advanced deep learning architectures (CNN, RNN, VAE, GAN).
Portfolio by: Matheus Ferreira TeixeiraEducation: Advanced Diploma in Software Engineering - AI, Centennial College (2022-2024)GPA: 4.45/4.5 (High Honors) | Ex-NASA Intern π
Key Achievements:
- π€ 15+ Production-Quality Implementations - From search algorithms to generative AI
- π 91.8% RΒ² Regression - Feature engineering for e-commerce prediction
- π¨ Generative Models - VAE latent space + DCGAN image synthesis
- π§ 88% Image Classification - CNN architecture for Fashion MNIST
- π¬ 92% NLP Accuracy - YouTube spam detection with Naive Bayes
- π¬ 11 Custom Neural Architectures - CNNs, RNNs, Autoencoders, VAE, GAN
Technical Depth:
- 5,000+ lines of Python code
- Complete ML pipelines (preprocessing β training β evaluation)
- Advanced techniques: Transfer learning, adversarial training, Ξ²-weighting
- Production-ready implementations with comprehensive documentation
π 00_AI_Fundamentals - Classical AI & Machine Learning
6 comprehensive projects covering core AI algorithms and supervised learning fundamentals.
Projects:
- Simple Reflex Agent - Environment simulation with percept-action mapping
- Graph Search Algorithms - BFS, UCS, Greedy, A* with visual tree exploration
- Linear Regression - Feature engineering (RΒ² 0.195 β 0.918 improvement)
- Logistic Regression - Binary classification (83% accuracy, threshold tuning)
- Neural Networks - Architecture experiments, overfitting analysis
- NLP Spam Detection - Text classification with TF-IDF + Naive Bayes (92% accuracy)
Skills: Search algorithms, regression, classification, neural networks, NLP, feature engineering, model evaluation
View AI Fundamentals Portfolio β
π§ 01_DeepLearning - Advanced Neural Architectures
4 cutting-edge projects implementing modern deep learning architectures for computer vision and generative AI.
Projects:
- CNN vs RNN Image Classification - Architecture comparison (88% vs 86.2% accuracy)
- Transfer Learning with Denoising Autoencoder - Semi-supervised learning (72.2% with 1,800 samples)
- Variational Autoencoder (VAE) - Probabilistic latent space (2D visualization, Ξ²-weighting)
- Deep Convolutional GAN (DCGAN) - Image synthesis (adversarial training, 35s GPU training)
Skills: CNNs, RNNs/LSTM, autoencoders, VAE, GANs, transfer learning, adversarial training, custom layers, convergence analysis
View Deep Learning Portfolio β
Programming & Computation:
- Python 3.10/3.11 - Primary language
- NumPy 1.23.5 - Numerical computing
- Pandas 1.5.3 - Data manipulation and analysis
Machine Learning:
- Scikit-learn - Classical ML algorithms and preprocessing
- Neurolab - Neural network experiments
Deep Learning:
- TensorFlow 2.10.0 / Keras - Deep neural networks
- TensorFlow Probability 0.18.0 - Probabilistic modeling (VAE)
Visualization & Analysis:
- Matplotlib 3.7.2 - Plotting and visualization
- Seaborn - Statistical visualization
- Graphviz/Pydot - Algorithm tree visualization
Natural Language Processing:
- NLTK - Text preprocessing
- CountVectorizer / TfidfTransformer - Feature extraction
Development Environment:
- Conda - Environment management
- Git/GitHub - Version control
- VS Code / Spyder - IDEs
- Google Colab - GPU training (T4 GPUs)
Algorithms:
- Graph search (BFS, UCS, Greedy, A*)
- Linear/Logistic regression with gradient descent
- Feedforward neural networks with backpropagation
- Convolutional Neural Networks (CNNs)
- Recurrent Neural Networks (RNNs/LSTM)
- Autoencoders (standard + denoising)
- Variational Autoencoders (VAE)
- Generative Adversarial Networks (GAN)
Training Strategies:
- Transfer learning with frozen encoder layers
- Unsupervised pre-training for semi-supervised learning
- Adversarial training with Nash equilibrium monitoring
- Ξ²-weighting for VAE loss balancing
- Custom early stopping (GAN-specific heuristics)
- Reparameterization trick for gradient flow
Data Engineering:
- Feature engineering and selection
- Categorical encoding (one-hot)
- Normalization (min-max, z-score)
- Text preprocessing (tokenization, TF-IDF)
- Train-validation-test splits
- Cross-validation
- Robotics navigation - Search algorithms for pathfinding
- Game AI - NPC behavior, decision-making agents
- Resource optimization - Scheduling and planning
- Image classification - Fashion item categorization (88% accuracy)
- Generative models - Synthetic image generation (VAE, GAN)
- Transfer learning - Pre-trained encoders for feature extraction
- Anomaly detection - Reconstruction error analysis (VAE)
- Customer analytics - Spending prediction (RΒ² 0.918)
- Survival prediction - Risk assessment (83% accuracy)
- Fraud detection - Anomaly detection techniques
- Marketing optimization - Customer lifetime value modeling
- Spam detection - YouTube comment classification (92% accuracy)
- Text classification - Sentiment analysis, content moderation
- Feature extraction - TF-IDF, Bag of Words
Linear Regression Project: Adding single feature ('Record' - purchase history) improved model from RΒ² 0.195 β 0.918 (+72% improvement)
Lesson: Domain knowledge and feature engineering can matter more than algorithm choice.
DCGAN Project: Identified Conv2DTranspose in discriminator violated DCGAN standard (Radford et al., 2015)
- Impact: Removed 211K parameters (50% reduction: 424K β 213K)
- Result: Restored adversarial balance (discriminator accuracy: 91% β 57%)
Lesson: Always validate architectures against established standards before implementation.
Autoencoder Project: Pre-trained encoder on 57,000 unlabeled images improved classifier trained on only 1,800 labeled samples
- Baseline: 70.33% validation accuracy
- Pre-trained: 72.17% validation accuracy
- Hardest class improvement: 8.8% β 32.4% (4x improvement)
Lesson: Unsupervised pre-training highly effective in low-data scenarios.
VAE Project: Principled Ξ² calculation via loss magnitude analysis (not trial-and-error)
- Problem: KL loss 100-1000x larger than reconstruction loss
- Solution: Ξ² = 0.001 targeting 15% KL contribution
- Result: Balanced training with sharp reconstructions + regularized latent space
Lesson: Analyze loss magnitudes to set hyperparameters systematically.
Neural Network Project: Simple architecture + 100 samples (test error: 0.0020) outperformed complex 2-layer + same data (test error: 0.0257)
Lesson: Focus on data quality before adding architectural complexity.
AI_Python_Projects/
βββ README.md # This file
β
βββ 00_AI_Fundamentals/ # Classical AI & ML (6 projects)
β βββ README.md # Fundamentals portfolio overview
β βββ 00_Agents/ # Simple Reflex Agent
β βββ 01_Search/ # BFS, UCS, Greedy, A*
β βββ 02_LinearRegression/ # E-commerce prediction
β βββ 03_LogisticRegression/ # Titanic classification
β βββ 04_NeuralNetwork/ # Architecture experiments
β βββ 05_NPL/ # YouTube spam detection
β
βββ 01_DeepLearning/ # Advanced DL (4 projects)
βββ README.md # Deep learning portfolio overview
βββ 00_CNN_RNN_ImageClassification/ # CNN vs RNN comparison
βββ 01_AutoencoderTransferLearning/ # Transfer learning pipeline
βββ 02_VAE_FashionMNIST/ # Variational autoencoder
βββ 03_ConvolutionalGAN/ # DCGAN image synthesis
Each project folder contains:
README.md- Comprehensive documentationenvironment.yml- Conda environment specificationrequirements.txt- pip dependencies- Source code (
.pyfiles) - Results (visualizations, plots)
- Python 3.10+ (3.11 for AI Fundamentals, 3.10 for Deep Learning)
- Conda (recommended) or pip
- GPU (optional, but recommended for Deep Learning projects)
1. Clone the repository:
git clone https://github.com/domvito55/AI_Python_Projects.git
cd AI_Python_Projects
2. Navigate to any project:
cd 00_AI_Fundamentals/02_LinearRegression
# or
cd 01_DeepLearning/00_CNN_RNN_ImageClassification
3. Create environment:
# Option A: Conda (recommended)
conda env create -f environment.yml
conda activate [environment-name]
# Option B: pip
python -m venv venv
source venv/bin/activate # Linux/Mac
# or
venv\Scripts\activate # Windows
pip install -r requirements.txt
4. Run the project:
python [main_script].py
| Project Category | CPU | GPU | Recommended |
|---|---|---|---|
| AI Fundamentals | 5-15 min | CPU sufficient | |
| CNN/RNN | 5-10 min | CPU acceptable | |
| Autoencoder | 10-15 min | CPU acceptable | |
| VAE | 30-35 min | GPU recommended | |
| GAN | Hours/Days | 35 seconds | GPU required |
Free GPU Access: Google Colab provides free T4 GPU (sufficient for all projects)
From Search Algorithms: A* optimal for pathfinding, Greedy fast but risky, BFS simple for unweighted graphs, UCS for cost optimization.
Takeaway: No universal "best" algorithm - match algorithm to problem characteristics.
From Logistic Regression: Threshold tuning creates precision-recall trade-offs
- 0.5 threshold: 83% accuracy
- 0.75 threshold: 81% accuracy, 100% recall
Takeaway: Business context determines optimal metrics beyond accuracy.
From Neural Networks: Training error 6.74Γ10β»βΆ looked excellent, test error 0.0826 revealed overfitting.
Takeaway: Always validate on separate test data; training error alone is deceiving.
From GAN: Oscillating losses are healthy in GANs (not training failure)
- Generator loss: 0.6-0.9
- Discriminator loss: 1.2-1.6
- Discriminator accuracy: 55-60% (balanced equilibrium)
Takeaway: GAN convergence requires multi-metric monitoring, not single loss minimization.
From VAE: Reparameterization trick (z = ΞΌ + Ο β Ξ΅) enables gradient flow through stochastic sampling.
Takeaway: Probabilistic modeling in neural networks requires careful gradient flow design.
Program: Advanced Diploma in Software Engineering - Artificial IntelligenceInstitution: Centennial College, Toronto, CanadaDuration: 2022-2024 (3-year intensive program)GPA: 4.45/4.5 (High Honors)
Key Courses:
- COMP 237: Artificial Intelligence (AI Fundamentals projects)
- COMP 263: Deep Learning (Deep Learning projects)
- Additional: Software Engineering, Data Structures, Algorithms
Industry Experience:
- 8-month Co-op: TD Bank (2x terms, 2023-2024)
- Multiple research projects with industry partners
- Ex-NASA Intern (2012, Python development for BOBCATT project)
Matheus Ferreira Teixeira
- π Portfolio: studentweb.cencol.ca/mferre39
- πΌ LinkedIn: linkedin.com/in/mathteixeira
- π¨βπ» GitHub: github.com/domvito55
- π§ Email: mathteixeira55@gmail.com
- π Location: BrasΓlia, Brazil (Open to remote opportunities)
Live Demos:
- Backend API: skillsdemo5.us-east-1.elasticbeanstalk.com
- Frontend App: skillsdemo-bucket.s3-website-us-east-1.amazonaws.com
- SpotSale API: spotsalebackend.uk.r.appspot.com
Available for:
- Full-time software engineering positions (AI/ML focus)
- Freelance AI/ML consulting projects
- Collaborative research opportunities
- Technical writing and documentation
Ex-NASA Intern (2012)
- Developed Python code for BOBCATT satellite project
- First professional experience demonstrating technical excellence from day one
Mensa Member
- Top 2.14% IQ globally
- Demonstrates analytical and problem-solving capabilities
Academic Excellence
- GPA 4.45/4.5 (High Honors)
- Consistent top performer throughout program
- Multiple research projects and industry collaborations
12+ Years Experience
- Engineering automation (2013-2017)
- Entrepreneurship and business management (2017-2023)
- Software engineering and AI/ML (2022-present)
Educational projects completed at Centennial College (2022-2024). Code available for learning purposes and portfolio demonstration with proper attribution.
Centennial College Faculty:
- Artificial Intelligence instructors
- Deep Learning instructors
- Co-op coordinators
Industry Partners:
- TD Bank (Co-op opportunities)
- Various research collaboration partners
Last Updated: November 2025
Comprehensive, production-ready implementations with professional documentation. Each project demonstrates end-to-end ML/DL pipeline development with real-world applicability.