I build complete AI systems — not just models. My work connects data pipelines, model training, explainability, retrieval, and backend APIs into structured, deployment-oriented workflows.

I am an AI engineering graduate from Jordan University of Science and Technology with a GPA of 3.72 / 4.00 — Excellent with Honors and 5× Dean's Honor List.

My strongest technical direction is building AI systems that are accurate, interpretable, and structured for real-world use — across Computer Vision, NLP Transformers, RAG pipelines, and Explainable AI.

I am currently deepening my practical work in Agentic AI — exploring tool-using LLM workflows, RAG-assisted reasoning chains, and structured decision-support agents.

RT-DETR · DINOv2 · U-Net · Attention U-Net · Xception · MobileNetV2 · DenseNet121 · AlexNet

CNN Techniques: Transfer Learning · Fine-tuning · Data Augmentation · Multi-class Classification · Semantic Segmentation · Object Detection

DistilBERT · MiniLM · ELECTRA-small · PubMedBERT

NLP Techniques: Fine-tuning · Token-level Analysis · Label Leakage Prevention · Large-scale Text Classification (~2.47M samples)

Prompt Engineering · Structured Outputs · Tool Use / Function Calling · LLM-assisted Report Synthesis · Agentic Workflows · Decision-Support Agents · Multi-step Pipelines

BM25 · Vector Embeddings · Semantic Search · Ranked Retrieval · Domain Knowledge Bases · Hybrid Retrieval (Dense + Sparse) · LLM-assisted Synthesis

BioMedCLIP · OpenCLIP · PubMedBERT · Vision-Language Embeddings · Medical Image Classification · U-Net Segmentation

LIME · Attention Maps · Token-level Explanations · Feature Attribution · Safety-critical Visual Inspection

Gradient Descent · SGD · Newton's Method · Genetic Algorithm (GA) · Particle Swarm Optimization (PSO) · Simulated Annealing

Applied to: Feature Selection · Hyperparameter Tuning · Wrapper Methods · Convergence Analysis

Data Cleaning & Preprocessing · Train/Val/Test Splitting · Class Imbalance Handling (SMOTE, Weighting) · Feature Engineering · Data Augmentation · Large-scale Batch Processing · Domain-specific Preprocessing (Medical, NLP, CV)

Exploratory Data Analysis (EDA) · Performance Benchmarking · Confusion Matrices · ROC/AUC Curves · Model Comparison Reporting

End-to-end AI Pipeline Architecture · Modular Backend Design · Deterministic + ML Hybrid Systems · Risk Assessment Logic · Structured Decision Engines · Safety-first System Design · Deployment-oriented Workflows

Pydantic · REST APIs · Structured JSON Outputs · Model Serving · API Documentation

FastAPI LightGBM SHAP FAISS RAG Docker Pydantic

My most complete AI engineering project. A production-style UAV pre-flight risk assessment backend that outputs structured GO / CAUTION / NO-GO decisions and a structured 11-section operational report covering ML risk, SHAP explanations, retrieved evidence, agent analysis, decision rationale, and recommended actions.

What makes it different: It does not rely on a single model. It combines deterministic safety validation, ML risk prediction, SHAP explainability, FAISS-based evidence retrieval, and a deterministic Decision Engine — with an LLM used only for report synthesis, not for decision-making.

Raw UAV Profile → Hard-Veto Safety Layer → ML Risk Prediction
→ SHAP Attribution → RAG Evidence Retrieval → Decision Engine
→ LLM Report Synthesis → Structured GO/CAUTION/NO-GO Report
 

PyTorch YOLOv8m RT-DETR DINOv2 Mini-Batch K-Means Ultralytics

Dense object detection on supermarket shelves with 170–230 products per image under occlusion and scale variation. Extended a single-class dataset into 91 visual product groups using DINOv2 embeddings and Mini-Batch K-Means clustering — without manual labeling.

PyTorch DistilBERT MiniLM ELECTRA-small LIME Gradio

Large-scale NLP classification on ~2.47M Reddit posts. Removed subreddit metadata to prevent label leakage. Fine-tuned three transformer models with token-level LIME explanations and a Gradio comparison interface.

Language-pattern classification prototype — not a medical diagnostic system.

TensorFlow Keras Xception MobileNetV2 DenseNet121 Grad-CAM LIME

Multi-class CNN classification across 42,900 images (Fire / Smoke / Non-Fire). Benchmarked three backbones with transfer learning. Used Grad-CAM and LIME to inspect model attention on safety-critical visual regions.

PyTorch BioMedCLIP U-Net OpenCLIP PubMedBERT Scikit-image

Medical imaging prototype combining foundation-model embeddings for classification and U-Net segmentation for lung-region mask prediction across 4 classes: COVID · Normal · Lung Opacity · Viral Pneumonia.

Medical image analysis prototype — not a clinical diagnostic system.

Python NumPy SciPy Matplotlib

Implemented and benchmarked 6 optimization algorithms from scratch — GD, SGD, Newton's Method, Genetic Algorithm, PSO, and Simulated Annealing — on a heart disease classification task. Used GA as a wrapper feature selector, reducing 20 features to 6 while maintaining competitive accuracy.

Driver Drowsiness Detection · Underwater Image Enhancement · Financial Fraud Detection · Facial Emotion Recognition @ ICICS 2025

• Driver Drowsiness Detection: AlexNet transfer learning with ~99%+ validation/test accuracy and Grad-CAM/LIME explanations.
• Underwater Image Enhancement: Attention U-Net with SSIM up to 0.7892 and Delta-E down to 0.02.
• Financial Fraud Detection: XGBoost achieved 99.4% accuracy on an imbalanced transaction dataset.
• Facial Emotion Recognition: ICICS 2025 research using Xception transfer learning; six-class FER model achieved 68.3% accuracy.

• 📄 Co-authored a Facial Emotion Recognition paper presented at ICICS 2025
• 🎓 AI Mentor — helped students with AI coursework and projects at JUST
• 🏆 Participated in AI, programming, and business hackathon environments

I do not treat AI as only model training. Every project follows a pattern:

Understand the domain
  → Clean data and prevent leakage
    → Train and compare models
      → Explain predictions (SHAP / LIME / Grad-CAM)
        → Design structured outputs
          → Build API or inference interface
            → Prepare for real-world use

This pattern is consistent across all my work — from UAV risk systems to medical imaging to NLP transformers.

• 🤖 Agentic AI — tool-using LLM workflows, RAG-assisted reasoning chains, and structured decision-support agents (actively learning and building)
• 🔗 Multi-step agent pipelines with memory, tool use, and grounded retrieval
• 📦 Deployment-oriented agent backends with FastAPI and structured outputs

Open to AI engineering opportunities in Computer Vision, NLP, RAG/LLM systems, and applied machine learning.