Widmark La Rosa

Fine-tuned YOLOv11 for real-time cube detection achieving 85%+ confidence. HSV-based color recognition pipeline reconstructs cube state, then Two-Phase solver computes near-optimal solutions in under 20 moves.

Full Transformer architecture in PyTorch — self-attention, multi-head attention, positional encoding, and layer normalization. Supports text generation, classification, and translation. Achieved ~80% validation accuracy.

Automated lost item metadata extraction using YOLOv8 + BLIP-2/Qwen2 for vision-language captioning. Generates categories, titles, and descriptions automatically — reducing manual input by ~60–70%.

Research investigating Vision Transformers vs CNNs under limited data fine-tuning conditions. Implements a HuggingFace ViT training pipeline alongside a PyTorch ResNet-18 CNN and compares both architectures on performance, convergence speed, and data efficiency.

Neural networks built from the ground up in pure Python — no frameworks. Implements forward propagation, backpropagation, and gradient descent manually to deeply understand how math translates directly into code and how deep learning models actually learn.

End-to-end ML pipeline that predicts customer churn using real telecom data. Covers exploratory data analysis, feature engineering, model training and evaluation, and ships a Streamlit app for interactive inference and visualization.