I design, evaluate, and deploy AI systems that move from research hypotheses
to reliable, secure, observable, and cost-aware production software.

I am an AI Researcher and Senior AI Engineer working at the intersection of scientific experimentation and production software engineering. My focus is building AI systems that are grounded in evidence, measurable under realistic evaluation, and maintainable after deployment.

My work spans LLM platforms, retrieval-augmented generation, agentic workflows, Arabic and multilingual AI, computer vision, OCR, speech systems, model optimization, secure backend services, and cloud delivery.

• Start from the problem, data, and measurable success criteria—not from a fashionable model.
• Separate prototypes from production systems through explicit reliability, security, latency, and cost requirements.
• Evaluate retrieval quality, hallucinations, edge cases, and failure modes before optimizing headline metrics.
• Use human review when uncertainty, operational risk, or domain impact is high.
• Build observable systems with reproducible experiments, versioned artifacts, and clear rollback paths.

Project-level outcomes reported from the systems described below.

Research intelligence platform for paper discovery, forensic analysis, claim verification, and paper-to-code generation.

• Built a GraphRAG-based workflow for parsing and analyzing technical papers across text, tables, figures, and OCR-extracted content.
• Developed a Scientific Code Forge that converts paper sections into modular Python and PyTorch project structures.
• Added claim-to-evidence checks that compare extracted claims with reported metrics, tables, and experimental results.
• Designed scientific auditing workflows using adversarial review, hype scoring, and consensus reporting.
• Mapped paper concepts to relevant GitHub implementations for implementation-oriented research.

Stack: Python, Asyncio, Streamlit, ChromaDB, GraphRAG, Docling, PyMuPDF, OCR, OpenAI, Anthropic, Gemini, Ollama
Repository: AI-Research-Intelligence-Agent

Enterprise support automation combining intent classification, RAG, multilingual interaction, and live operations.

• Automated 60–80% of customer inquiries through intent classification and retrieval-grounded response generation.
• Delivered English and Arabic interaction with automatic RTL handling.
• Built workflows for agent takeover, case management, orders, staff administration, and operational monitoring.
• Implemented JWT authentication with HTTP-only cookies and role-based access control.
• Added analytics for latency, error rates, intent distribution, traffic, and operational quality.

Stack: Next.js, React, TypeScript, Node.js, Cloud Run, Vertex AI Gemini, Firestore, BigQuery, Docker
Demo: NexusAI · Repository: nexus-ai

Low-latency multilingual voice assistant with Arabic dialect classification.

• Combined Whisper v3, a fine-tuned MARBERTv2 classifier, and downstream LLM reasoning in a modular inference pipeline.
• Targeted sub-500 ms response time for short utterances.
• Improved dialect-recognition F1 by 12% over a Whisper-only baseline.
• Built real-time communication using Flask, React, REST APIs, and WebSockets.
• Orchestrated retries, queues, events, and human review using n8n and Elsa Workflows.

Stack: Whisper, MARBERTv2, Flask, React, WebSockets, n8n, Elsa Workflows
Repository: AIVoiceAsisstent

Deployment-oriented Arabic character recognition with model compression and CPU optimization.

• Trained a CNN on 120,000 handwritten samples across 28 classes.
• Achieved 85% top-1 accuracy using preprocessing, augmentation, and dropout regularization.
• Applied quantization-aware training and ONNX conversion.
• Reduced model size by 59% and improved CPU inference speed by 2.3×.
• Exposed the model through a Flask API for public use.

Stack: TensorFlow, Flask, ONNX, PythonAnywhere
Demo: LearnWithUs · Repository: LearnWithUs

• Scientific AI agents, paper analysis, and paper-to-code systems
• GraphRAG, long-context retrieval, reranking, and knowledge-grounded reasoning
• LLM evaluation, hallucination detection, claim verification, and reliability engineering
• Arabic NLP, dialect-aware intelligence, speech systems, and multilingual assistants
• OCR, document intelligence, and multimodal verification pipelines
• Efficient inference, quantization, ONNX, model serving, and cost-aware deployment

Methods: LLMs, RAG, GraphRAG, embeddings, reranking, structured outputs, prompt evaluation, guardrails, OCR, speech AI, computer vision, fine-tuning, PEFT/LoRA, RLHF-style workflows, quantization, and ONNX.

Engineering: ASP.NET Core, Entity Framework, REST APIs, WebSockets, React, FastAPI, Flask, Docker, IIS, CI/CD, OpenTelemetry, Prometheus, Grafana, MLflow, and Weights & Biases.

Problem framing
    → hypothesis and success criteria
    → data and evidence audit
    → reproducible prototype
    → evaluation and failure analysis
    → security, latency, cost, and reliability hardening
    → monitored deployment
    → evidence-driven iteration

Research-backed AI. Production-grade engineering. Systems designed to be evaluated, operated, and trusted.

LinkedIn · Email · Portfolio · Repositories