Hamza – MCP, PyTorch, LLM evaluation

AI-powered evaluation product for very small businesses, designed to benchmark how leading language models handle real-world operational, customer-support, marketing, and decision-making tasks. The product compares model outputs across accuracy, reasoning, instruction-following, safety, and practical usefulness, turning structured evaluations into actionable insights for selecting and improving AI solutions.

• Evaluated and compared responses from leading AI models for very small business use cases;
• Assessed model quality, ranked outputs, and performed detailed performance analysis;
• Benchmarked frontier LLM outputs for factual accuracy, reasoning quality, instruction-following, and safety alignment;
• Identified logical inconsistencies and edge-case errors and produced structured feedback for iterative model improvement;
• Applied consistent guideline-driven judgment across high volumes of evaluations to maintain QA calibration.

A machine learning regression solution for predicting freight rates from historical shipment data. The project processed 48,000 historical freight records containing pickup and delivery locations, geographic coordinates, distance, equipment type, weight, date, market conditions, and quote signals. I developed an end-to-end machine learning pipeline covering data exploration, feature engineering, preprocessing, model training, hyperparameter tuning, chronological validation, and prediction generation for new loads. The solution was designed to support data-driven freight pricing and forecasting.

• Built an end-to-end machine learning pipeline for freight rate prediction;
• Performed exploratory data analysis to identify missing values, distributions, correlations, and relationships between shipment features and freight rates;
• Engineered date-based features including year, month, day, and day of week;
• Implemented preprocessing pipelines for numerical and categorical features using median imputation, most-frequent imputation, and OneHotEncoder;
• Used a chronological train/validation split to simulate real-world future prediction and reduce temporal data leakage;
• Trained and evaluated a RandomForestRegressor for nonlinear freight-rate prediction;
• Tuned model hyperparameters using GridSearchCV and selected an optimized configuration with 200 estimators and a minimum leaf size of 2;
• Generated predictions for the validation dataset and a December 2025 forecasting scenario;
• Created reusable Python scripts for data exploration, feature engineering, model training, scoring, and prediction generation;
• Documented the complete modeling workflow and validation methodology.

An AI-powered multidisciplinary decision-support platform that uses specialized AI agents to collaboratively analyze patient and athlete cases. The system combined kinesiology, sports psychology, biomechanics, data analysis, and safety expertise to evaluate movement, injury, rehabilitation, and return-to-play scenarios. Designed a multi-agent workflow where agents exchanged findings and reasoning before producing a unified, safety-focused recommendation.

• Spearheaded the architecture of a trilingual RAG platform for movement-science and neuromuscular research queries;
• Designed the end-to-end retrieval and reasoning pipeline from the ground up;
• Led the design of Random Forest and LSTM models on a longitudinal biometric dataset of 5,500+ labelled records;
• Directed the translation of applied kinesiology and movement-science research into a working AI analytics product.

AI product portfolio delivering intelligent automation and decision-support tools for business, healthcare, education, and agriculture. The products combined predictive modeling, classification, anomaly detection, multilingual document intelligence, and automated report generation to transform complex data into practical insights.

• Architected and deployed ML and DL pipelines for classification, prediction, and anomaly detection across health-tech, EdTech, agriculture, and business domains;
• Designed Agentic AI workflows and Advanced RAG pipelines for Arabic, French, and English document analysis and report generation;
• Engineered feature-engineering pipelines for tabular and time-series data and improved model accuracy by up to 18% across selected projects;
• Shipped ML services to AWS SageMaker via Docker and CI/CD and added model monitoring and AI safety controls, reducing deployment cycle time by approximately 40%.

An AI-powered mental health assistant providing 24/7 emotional support, distress detection, and wellness guidance. Built with FastAPI and NLP to offer a compassionate, confidential digital companion.

• Designed and developed an AI-powered mental health assistant for 24/7 conversational support and wellness guidance;
• Built and maintained backend services using Python and FastAPI;
• Integrated NLP and Claude LLM capabilities to generate contextual, empathetic responses;
• Designed conversational workflows for personalized wellness guidance and user engagement;
• Implemented distress-signal detection and safety-oriented response flows;
• Prioritized privacy, reliability, and responsible AI practices across the product experience.

An advanced biomechanical analysis system designed to analyze human movement and estimate physical performance and impact metrics using physics-based calculations and physiological data. The system calculates ground reaction forces, impact forces, energy expenditure, and heart-rate variability metrics across different movement surfaces. It also supports automated PDF report generation and a Streamlit-based interface for analyzing movement-related data and preparing the system for future video-based analysis with MediaPipe

• Built an end-to-end Python-based biomechanical analysis system combining machine learning concepts with physics-based calculations;
• Implemented ground reaction force calculations in both pounds-force and Newtons;
• Developed impact force and G-force analysis for movement assessment;
• Implemented energy expenditure calculations across different training surfaces;
• Developed HRV-based recovery analysis functionality;
• Built a Streamlit interface for interacting with the analysis system;
• Implemented automated PDF report generation for analysis results;
• Structured the application into modular components covering biomechanics, physics, ML models, database, protocol, frontend, and reporting;
• Designed the architecture to support future video-based movement analysis using MediaPipe;
• Tested and organized the system into reusable Python modules.