AI & Machine Learning Projects

Overview

Developed a model forecasting sales for 6 restaurants (10k+ records) with a 98.41% accuracy R2 and outperforming an already high naive baseline of 97.7%. Used GridSearch for hyperparameters and testing algorithms (RNNs, classical algorithms, Neural Networks), and XGBoost was chosen based on RMSE, R2, and naive baseline comparison.

Key Features

• Custom CNN architecture with 5 convolutional layers optimized for multi-class classification
• Advanced data augmentation pipeline (rotation, flipping, scaling) to improve model generalization
• Real-time image classification with sub-second inference times (avg. 0.3s per image)
• Interactive web interface built with Flask for easy testing and demonstration
• Model versioning and experiment tracking with TensorBoard

Technical Stack

Challenges & Solutions

The main challenge was dealing with class imbalance in the dataset, where some categories had 3x more samples than others. I addressed this by implementing weighted loss functions and applying SMOTE (Synthetic Minority Over-sampling Technique) to balance the training data. This improved minority class accuracy by 18%.

Results & Impact

The final model achieved 94% overall accuracy with F1-scores above 0.90 for all classes. The system processes images 5x faster than the baseline approach and has been deployed as a demonstration project showcasing practical computer vision applications.

Overview

Developed a deep learning model predicting the % that an new client will default on a loan (10k+ records). Target data was heavily imbalanced (2.91 Mean ABS Skewness) so I mainly focused on PR-AUC and our Recall based off of F1 Optimal Threshold. Our deep learning model caught 44% of actual defaults (Recall), and PRC-AUC of 0.286 wasa 79% improvement over randomly flagging loan applicants based on our data's historical default rate of 16%--resulting in a savings of $4.1M. F1 Optimal Threshold was an aggressively low 0.19 which considered both Precision and Recall but weighed Recall more heavily so we could focus more on the minority class (defaults).

Key Features

• Context-aware responses using attention mechanisms and BERT embeddings
• Real-time sentiment analysis of user inputs with 89% accuracy
• Multi-turn conversation handling with sliding window memory (up to 10 previous exchanges)
• Integration with popular messaging platforms (Slack, Discord)
• Customizable personality traits and response styles
• RESTful API for easy integration into existing applications

Technical Stack

Challenges & Solutions

Managing conversation context over long dialogues was challenging, as storing full conversation history quickly consumed memory. I implemented a sliding window approach combined with Redis caching to efficiently store and retrieve conversation history without sacrificing response time. This reduced memory usage by 60% while maintaining context quality.

Results & Impact

The chatbot successfully handles complex multi-turn conversations with an average response time of 1.2 seconds. User testing showed 85% satisfaction rate with response relevance and naturalness. The system is currently being enhanced with retrieval-augmented generation (RAG) for domain-specific knowledge.

Overview

Built a Convoluted Neural Network (CNN) off a ResNet50v2 backbone to detect multiple vehicles and their category type from images with an accuracy of 87% and an IoU of 88%

Key Features

• Drag-and-drop CSV/Excel file upload with automatic schema detection
• Automated data preprocessing and feature engineering (handling missing values, encoding, scaling)
• Multiple ML algorithm selection (Random Forest, XGBoost, LSTM, Linear Regression)
• Interactive visualizations with Chart.js and D3.js (scatter plots, time series, correlation matrices)
• Model performance metrics and comparison tools (RMSE, MAE, R², confusion matrices)
• Export predictions and trained models for future use
• Responsive design optimized for desktop and tablet devices

Technical Stack

Challenges & Solutions

Handling large datasets in the browser was initially problematic, causing slowdowns and memory issues for files over 50MB. I solved this by implementing serverless processing with AWS Lambda for heavy computations, while using React with virtualized lists for a responsive and smooth user interface. This architectural decision reduced client-side memory usage by 80% and improved processing speed for large datasets.

Results & Impact

The dashboard successfully processes datasets up to 1GB in size and generates predictions within 30 seconds for most common use cases. The application has been used for exploratory data analysis and quick prototyping of ML models, reducing the time from data to insights by approximately 70% compared to traditional coding approaches.