Youssef Naggar

Email Fraud Detection Pipeline

Tri-Stream Stacking Ensemble + Deep Autoencoder

F2: 0.8842 | 90.5% Recall | ROC-AUC: 0.999

An end-to-end fraud classification architecture built on 447K Enron emails to resolve an extreme 192:1 class imbalance. Incorporates a custom PyTorch Deep Autoencoder to compress 384-dimensional MiniLM embeddings into 64-dimensional latents and derive unsupervised anomaly reconstruction error signals, stacked with a Tri-Stream Level-1 GBDT ensemble and an ElasticNet meta-learner.

System Specifications:

PyTorch Autoencoder: 384-d MiniLM compressed into 64-d latents + unsupervised anomaly MSE reconstruction signal.
Tri-Stream Level-1 Models: Stream A (GPU XGBoost), Stream B (CatBoost with native text dictionaries), Stream C (LightGBM on TF-IDF n-grams).
Level-2 Stacking Meta-Learner: ElasticNet meta-learner in logit space with 12 skip-connection domain features (VADER polarity, stylometry, sender velocity).
Dynamic Threshold Tuning: tau* = 0.050 achieving 0.8842 F2-score, 90.5% recall, 80.8% precision, and 0.999 ROC-AUC.

PyTorch XGBoost CatBoost LightGBM Autoencoders MiniLM Optuna

View GitHub Repository

Weather Wizard 3000

Vectorless RAG + LiteLLM Multi-Modal Engine

LiteLLM + Vectorless RAG

A personalized AI outfit recommendation assistant leveraging Vectorless RAG and LiteLLM for SDK-agnostic model routing. Synthesizes 5-day real-time OpenWeatherMap forecasts with structured multimodal closet inventories and user preferences to generate grounded, weather-appropriate wardrobe suggestions with virtual avatar rendering.

System Specifications:

Multimodal Closet Synthesizer: Processes clothing photos via Vision LLM into structured metadata stored in closet.json.
Grounded LLM Weather Wizard: Filters real-time weather forecasts and dynamically matches seasonal attributes with available wardrobe items.
Virtual Avatar Drawer: Generates personalized visual demo previews combining user avatar and recommended outfit items.
Employs LiteLLM for seamless SDK-agnostic model switching and automated IP/GPS geolocation detection.

Python LiteLLM Vectorless RAG OpenWeatherMap Multimodal Vision

View GitHub Repository

Dog Breed Classifier (ConvNeXt-V2)

Computer Vision Fine-Tuning Pipeline

90.5% Test Accuracy (120 Breeds)

A deep learning computer vision pipeline trained on the Stanford Dogs dataset across 120 breeds. Features automated bounding-box image preprocessing to isolate subjects, MixUp data augmentation for cross-class generalization, and fine-tuning of ConvNeXt-V2 in PyTorch with customized learning rate scheduling.

System Specifications:

Architecture: ConvNeXt-V2 with 7x7 depthwise convs, inverted bottleneck blocks, and Global Response Normalization (GRN).
Preprocessing: Automated bounding-box image cropping to normalize training data and eliminate background noise.
Generalization: MixUp augmentation and cosine annealing learning rate scheduler in PyTorch.

PyTorch ConvNeXt-V2 MixUp Computer Vision

View GitHub Repository

MicroGrad - Custom Neural Network Engine

Vectorized Backpropagation Engine + GPU Acceleration

84% EMNIST Match (CuPy GPU)

A ground-up implementation of a neural network engine in Python with vectorized backpropagation, Momentum SGD, and GPU acceleration via CuPy. Matches PyTorch benchmark performance (84% test accuracy on EMNIST) without relying on external deep learning frameworks.

System Specifications:

Forward & Backward Engine: Pure Python matrix calculus with vectorized derivatives for ReLU, Sigmoid, and Tanh.
Optimization: Momentum SGD optimizer with He and Xavier weight initialization schemes.
Hardware Acceleration: CuPy integration for high-throughput GPU matrix operations with automatic CPU fallback.

Python NumPy CuPy GPU Matrix Calculus

View GitHub Repository

UltraTube - Asynchronous TUI Media Downloader

Textual Terminal UI + Pytest Suite

Textual TUI + AsyncIO

A high-performance, multi-tab asynchronous Terminal User Interface (TUI) application built with Python's Textual framework and yt-dlp. Features parallel queue processing, live download progress bars, metadata processing (merging audio tracks, thumbnails, and dual subtitles), and comprehensive test coverage following Dependency Inversion.

Python Textual (TUI) yt-dlp FFmpeg Pytest

View GitHub Repository

Wordle Combination Finder

Information Gain Optimization Algorithm

+8% Heuristic Superiority

An algorithmic search system identifying optimal 4-word starting combinations for Wordle using letter position frequency analysis, a custom Trie data structure, and parallel backtracking with multiprocessing.

Key Finding:

Optimal 4-word combination: ['HOWFS', 'IXTLE', 'JUMBY', 'KRANG'] with a score of 2659, outperforming commonly cited combinations by 8%.

Python Algorithms Trie Multiprocessing

View GitHub Repository

Monster Wrangler Game

2D Arcade Game Engine

Pygame OOP

A fast-paced 2D arcade game built with Pygame featuring object-oriented game state management, sprite animation, dynamic collision handling, and progressive difficulty.

Python Pygame OOP

View GitHub Repository