Machine Learning Projects

This repository showcases a growing collection of my machine learning projects, ranging from classic ML algorithms to advanced deep learning and applied research. These examples highlight my experience with classification, regression, optimization, neural networks, reinforcement learning, and bioinformatics. Projects are implemented using tools like NumPy, scikit-learn, PyTorch, LangChain, and transformers.

In addition to individual projects, this repository also references collaborative work such as SmartStudy (academic habit optimizer using GPT-4 and TabPFN) and Cognify AI (an AI-powered educational assistant for kids built with DeepPiXEL and UTESCA).

---

🧠 Projects Included

1. Pocket Algorithm & Linear Regression

• Implemented binary classification using the Pocket Algorithm on the IRIS dataset
• Linear regression with mean squared error on the Diabetes dataset
• Compared performance with scikit-learn baselines

2. Gradient Descent & Multiclass Logistic Regression

• Implemented four variants of stochastic gradient descent: SGD, Heavy-ball, Nesterov, and Adam
• Built multiclass logistic regression from scratch with one-hot encoding and softmax
• Compared optimizer performance on IRIS and Digits datasets

3. Feedforward Neural Networks (NumPy + PyTorch)

• Implemented all layers of a 2-layer neural network using NumPy
• Trained and tuned on the IRIS dataset using cross-entropy loss
• Built and trained a PyTorch MLP on FashionMNIST using nn.Linear, ReLU, and CrossEntropyLoss

4. Recurrent Neural Networks & Reinforcement Learning

• Vanilla RNN and LSTM for the Adding Problem (sequence regression)
• Value Iteration and Q-Learning agents applied to GridWorld, Crawler, and Pacman
• Approximate Q-Learning using feature-based representations

5. RL-Based Medication Adjustment for Speech Disfluency (Research Project)

• Conducted as part of a research project under the supervision of Prof. Michela Rughazroy, in collaboration with UofT Engineering students
• Helped with data collection, research and development a reinforcement learning-based system aimed at adjusting medication to reduce speech disfluency
• Published Paper

6. Stress Detection with EEG & ECG (STEM Fellowship AI Challenge 2024)

• Built a multi-perceptron neural network to classify stress levels (normal, high, low) using EEG and ECG signals
• Processed signals from 40 participants during mental arithmetic tasks
• Achieved competitive performance (~62% accuracy) compared to models like SVM, RF, and AdaBoost
• Project published on the STEM Fellowship website
• GitHub Repository
• Published Paper

7. Predicting Disease-Causing Mutations (Bioinformatics Hackathon 2024)

• Developed during a bioinformatics hackathon in Jan 2024
• Built a machine learning model to predict whether a genetic sequence contains mutations related to Alzheimer’s disease, with a focus on women's health
• Used genomic mutation data from DSS NIAGADS and GenomeKit, filtered based on p-values < 0.05
• Built features using biological mutation transformations, trained a Random Forest classifier combined with Borzoi from gReLU for mutation effect prediction
• Challenges involved unfamiliarity with genomics and time-limited setup; overcame these through collaboration and rapid research
• Inspired by the disproportionate impact of Alzheimer’s on women and the shortage of genetic prediction tools
• Future improvements include increasing accessibility for clinicians and refining mutation localization
• 🔗 View Project Repository

8. SmartStudy: Personalized Academic Recommendation System (ECE324-Machine Intelligence, Software and Neural Networks Course Project)

• A collaborative project focused on predicting and improving student GPA based on lifestyle and study patterns
• Models used: TabPFN, CatBoost, XGBoost, MLP, TabNet, 1D CNN, and stacked ensembles
• Used Bayesian Optimization to recommend behavior changes that match students’ GPA goals
• Integrated K-Nearest Neighbors for matching similar students and GPT-4 to convert outputs into academic advice
• Includes a Gradio web app for user interaction
• 🔗 View Project Repository

9. Cognify AI: Adaptive Learning Assistant for Kids (UTESCA Project)

• Built as part of the UTESCA x DeepPiXEL consulting initiative as ML Developer
• A GenAI solution for personalized education targeting children aged 8–12
• Won People' Choice Award at URESCA CEP 2025 event
• I contributed as an ML developer focusing on integrating multiple AI models
• Features include:
  • Text chat using LangChain + GPT-3.5
  • Voice chat via Whisper + GPT + gTTS
  • Object recognition via Vision Transformer (ViT)
  • Handwriting-to-math solver using EasyOCR and TrOCR
  • Real-time transcription and TTS
  • Adaptive learning progress tracking
  • Security features like encryption and authentication
• The codebase is private due to project confidentiality.

---

🚀 Coming Soon

This repository will be expanded with more advanced machine learning projects.