MAHMOUDREFAEY
Turning research-driven AI into scalable, real-world products.
About
I'm an AI Engineer building LLM applications (RAG, agents, Text-to-SQL), including Arabic document and OCR pipelines. At Global Brands Group (GBG), I helped raise an Arabic banking RAG system's faithfulness from 79% to 97% and co-developed Chinhook, an open-source Text-to-SQL agent. I also fine-tuned Gemma 4 E2B for chart reasoning and built OrbitTalent, an LLM-based CV screening platform. I thrive at the intersection of deep learning research and practical software engineering — building systems that don't just work in notebooks, but scale in the real world.
Military Status: Exempted ✓{
"name": "Mahmoud Refaey",
"role": "AI Engineer",
"location": "Cairo, Egypt",
"status": "Open to Opportunities",
"languages": ["Arabic (native)", "English (fluent)"],
"education": "B.Sc. Artificial Intelligence (Honors)",
"interests": ["LLM Applications", "RAG & Agents", "Arabic NLP & OCR"]
}Technical Arsenal
Featured Projects
Chinhook
Open-Source Text-to-SQL Agent
Open-source Text-to-SQL agent, co-developed as an AI Engineer and sponsored by GBG. Designed the LangGraph agent workflow: query rewriting, schema retrieval from Qdrant, SQL generation, and verification before any answer is returned. Restructured the codebase for multi-tenancy, so each user connects their own database and OpenAI-compatible model without credentials being stored. Deployed the internal version on Azure with Docker and Azure OpenAI.
OrbitTalent
LLM-Assisted CV Screening Platform
Multi-tenant applicant-tracking platform (sole developer) that scores CVs against a job description and returns an explainable ranked shortlist: fit score, job-match %, and matched/missing skills. A four-tier scoring cascade (keyword gate, from-scratch BM25, low-cost LLM, deeper LLM) stops at the first confident tier, so paid model calls run only where needed, and token cost is logged per call. Tenants are isolated at the query layer (cross-tenant requests return 404), with JWT sessions and per-IP rate limiting on auth endpoints. Deployed on Vercel and AWS EC2 over HTTPS.
Clarivis-R1
Chart Reasoning Vision-Language Model
Fine-tuned Gemma 4 E2B with LoRA on vision and language layers, using 300K chart-reasoning samples from IBM Granite's ChartNet, so the model reasons step by step before each answer. Built the training pipeline for 300K chart images on a single AWS SageMaker GPU (about 5 days), resolving dataset serialization and GPU memory issues. Reached 51.64% relaxed accuracy on the 2,500-question ChartQA test set, and published the model with a card covering setup, evaluation, and limitations.
SpendHiST
AI-powered receipt management & spending analytics
OCR and LLM extraction pipeline that turns English and Arabic receipts into structured data, with 95% extraction accuracy on English receipts and 83% on Arabic. A LangGraph multi-agent backend routes questions between SQL retrieval and receipt-analytics agents, with JWT authentication and rate limiting.
AraSpellX
Arabic spelling correction with Transformer seq2seq
Independent Transformer encoder-decoder implementation for Arabic spelling correction, inspired by the AraSpell research paper. Generates synthetic spelling errors (insertions, deletions, substitutions) from clean Arabic text and evaluates with CER and WER. Modernized the training pipeline with PyTorch 2.3 scaled-dot-product attention, BF16 mixed precision, gradient accumulation, and a pre-tokenized disk cache, cutting epoch time on 6.9M training samples from ~7 days to ~3 hours (56x) and peak GPU memory from 10.6 GB to 1.4 GB.
NexaOS
AI-Powered Desktop Environment
AI-powered web desktop environment with intelligent document editors, custom file formats (.nd, .np, .ndf), and real-time collaboration via Socket.io. Built an assistant agent for the marketing team, a RAG chatbot for questions about the system's content and documents, and in-app AI features such as AI text generation in a Notepad-style editor.
Interactive Multi-PDF Chat
RAG pipeline for semantic PDF search
Conversational AI system using FAISS + TinyLlama/Phi-2 + Sentence-Transformers for semantic search over multiple PDF documents. Deployed on Streamlit Cloud.
AI Churn Prediction Platform
Graduation Project — Telecom customer churn prediction
Stacking ensemble of XGBoost, LightGBM, and Random Forest with 84.5% PR-AUC, served through Django REST prediction APIs with automated preprocessing, prediction history, and export. Led a team of 6 through development, documentation, and deployment.
Land Cover Classification
ResNet50-based satellite image classification (EuroSAT)
Deep learning app for real-time EuroSAT land-cover classification across 10 categories with confidence scoring, visual analytics, and a modular Streamlit interface integrating TensorFlow, OpenCV, and Plotly.
Energy Consumption Forecasting
Time-series ML pipeline for household electricity prediction
End-to-end ML pipeline forecasting hourly electricity usage on the UCI dataset (2M+ records). Engineered lag features & time-based attributes. XGBoost achieved R²=0.58, MAE=0.32; Prophet enabled interpretable seasonal forecasting.
Experience
Global Brands Group (GBG)
AI Engineer Intern(Hybrid)
Worked on most stages of an Arabic RAG chatbot over a regional bank's internal PDFs in Azure Blob Storage: EasyOCR ingestion, cleanup, chunking, hybrid retrieval (BM25 over normalized Arabic text plus text-embedding-3-large in Azure AI Search, merged by weighted reciprocal rank fusion, then cross-encoder reranking), query rewriting (multi-query, HyDE, decomposition), routing, and evaluation. Helped raise faithfulness from 79% to 97% and correctness from 61% to 85% (DeepEval, LLM-as-judge). Also co-developed Chinhook, an open-source LangGraph Text-to-SQL agent, restructured it for multi-tenancy, and deployed the internal version on Azure.
Beetleware
AI Engineer Intern(Remote)
Worked on NexaOS, an AI-powered web desktop application, across AI and full-stack development. Built an assistant agent to support the marketing team and a RAG chatbot for questions about the system's content and documents, and added AI features inside NexaOS apps, such as AI text generation in a Notepad-style editor. Also fine-tuned models, built data pipelines, and contributed to open-source projects, shipping through CI/CD.
Digital Egypt Pioneers Initiative (DEPI)
Data Science Trainee(Hybrid)
Built and evaluated machine learning models for classification and regression tasks using Python, Pandas, scikit-learn, and TensorFlow. Worked on industry-simulation projects covering end-to-end data science workflows — from data cleaning and EDA to model selection and evaluation. Mentored peers in core ML and data analysis concepts.
Education
Egyptian Russian University (ERU)
B.Sc. in Artificial Intelligence
Grade: Very Good (Honors)
Graduation Project
Telecom Churn Prediction Platform (Team Lead, team of 6)
Key Courses
Certifications
IBM AI & Data Science
DEPI / MCIT
Deep Learning with PyTorch
Mahara Tech (ITI)
Machine Learning Specialization
Coursera / DeepLearning.AI & Stanford
Certified Python Data Associate
DataCamp
Developing Applications with LangChain
DataCamp
Multi-Agent Systems with LangGraph
DataCamp
AWS Cloud Practitioner Essentials
Amazon Web Services
ML & DL Training
Zewail City
Research & Publications
Customer Churn in Telecom: Predictive Modeling for Enhanced Retention Strategies
Egyptian Russian University
Stacking ensemble approach combining XGBoost, LightGBM, and Random Forest with a Django-React full-stack analytics platform.
Read PaperA Study of Generative Approaches for Balancing Imbalanced Data: SMOTE, GANs, and LLMs
Published on ResearchGate
Compared SMOTE, GANs (CTGAN/TVAE), and LLMs for fraud detection; traditional sampling proved more stable than the generative approaches.
Read PaperContact
Let's Build
Something Great
I'm always open to discussing new projects, creative ideas, or opportunities to be part of something amazing.