Engineering that ships and scales

Architected a production-grade multimodal reasoning pipeline combining Qwen2.5-VL, OCR, and Zero-Shot Chain-of-Thought prompting. Achieved 93.96% accuracy on the ChartQA benchmark — outperforming state-of-the-art 72B parameter models at a tenth of the compute cost on an RTX 3090. Published at IEEE CAI 2026, sponsored by Microsoft.

Designed and deployed a scalable data ingestion and tagging pipeline for a major university's digital library. Integrated Florence-2, Tesseract OCR, YOLO, and NLP tooling to automate metadata extraction, visual classification, and semantic indexing across hundreds of thousands of documents.

Built a full-stack mobile application that combines Gemini API for intelligent recipe generation with a YOLO-powered computer vision layer that automatically detects ingredients from photos. The system delivers personalized, context-aware culinary recommendations in real time.

Engineered an AI-driven precision agriculture platform using ensemble machine learning models to predict soil moisture and optimize irrigation scheduling. Reduced water usage while maintaining crop yield through predictive water management and sensor fusion.

Delivered an end-to-end ETL and forecasting system for one of Egypt's largest banking institutions. Combined Prophet and SARIMAX time-series models to predict ATM cash fluctuations, directly reducing operational costs and replenishment cycles. Ranked #1 out of 1,000+ interns for project impact.

Built a financial forecasting engine that fuses LSTM and Transformer architectures with multi-head attention mechanisms and Quantile Regression to predict price volatility. Integrated TA-Lib for technical indicator preprocessing, enabling robust uncertainty-aware predictions across multiple market regimes.

Led the AI and navigation subsystem of a fully autonomous self-driving golf cart built from the ground up. Developed deep learning models for real-time drivable area segmentation, obstacle detection, and sensor fusion — integrating computer vision with onboard navigation hardware for end-to-end autonomy.

Developed a robust illumination-aware image segmentation system for outdoor environments using color space conversion, adaptive thresholding, texture entropy analysis, and morphological operations — achieving reliable shadow/non-shadow classification under variable lighting conditions.

Built a DSP and ML pipeline for classifying neural activity patterns from EEG signals associated with motor imagery tasks. Applied signal preprocessing, feature extraction, and a KNN classifier to distinguish between cognitive states with high accuracy.

Architected a full-stack offline-first PWA serving multi-role users (servants, parents, congregants) with real-time sync via Firebase RTDB, Edge-runtime API routes, and an IndexedDB-backed conflict-resolution queue for seamless offline mutations. Features role-based access control, structured attendance tracking, aggregated analytics dashboards, Zod validation, Sentry monitoring, i18n (Arabic/English), rate-limiting via Upstash Redis, and automated CI/CD — deployed on Cloudflare Edge with sub-second response times.

Engineered a full end-to-end data intelligence platform including automated web scraping, structured database management, and a polished desktop GUI. Built to aggregate, normalize, and surface labor market insights from Egypt's leading job portal.

Designed and trained a custom transformer encoder from scratch — embedding layers, multi-head self-attention, feedforward blocks — for large-scale sentiment classification. Achieved competitive accuracy on the IMDB benchmark without relying on pre-trained language model weights.

Built an unsupervised topic modelling system using Latent Dirichlet Allocation to extract and surface dominant themes across large document corpora. Designed for research and content analytics workflows requiring interpretable, scalable semantic clustering.

Developed and benchmarked a suite of NLP classifiers — LinearSVC, Naïve Bayes, Decision Trees, and MLP — for large-scale social media sentiment analysis. Produced a rigorous comparative evaluation framework for production model selection.

Conducted an original empirical study on the interplay between neural network depth and width across diverse application domains. Findings were presented at the EURECA 2025 conference — contributing to the research community's understanding of architecture design tradeoffs.

Designed and benchmarked a comprehensive computer vision pipeline for emotion classification, evaluating seven architectures including VGG16/19, ResNet50/101, InceptionResNetV2, Xception, and a custom CNN — with full SHAP explainability analysis.

Trained character-level generative language models on Shakespeare's complete works using LSTM, GRU, and Bidirectional RNN architectures — exploring sequence modelling, creative text generation, and the qualitative differences between recurrent topologies.