Built and shipped at scale.

Dubai, United Arab Emirates

Agentic AI and GenAI systems in production for the largest travel company in the Middle East, alongside revenue forecasting and MLOps platform work.

• Re-architected a single-call LLM extraction system for airline booking emails into a multi-step agentic workflow (LangGraph, LangChain) with structured output validation, self-correction loops, and fallback handling, significantly improving extraction accuracy and reliability.
• Built an agentic voice AI system (Retell AI) with LLM-driven dialogue management and tool calling to assist customer service, optimizing response quality and call-handling latency for real-time customer interactions.
• Designed a post-call GenAI pipeline using LLM-based summarization, classification, and entity extraction to convert raw call transcripts into structured records automatically.
• Deployed a time-series forecasting system for flight add-on revenue, enabling revenue optimization decisions that increased add-on revenue by 55%.
• Developed an XGBoost financial-forecasting model for budgeting that cut prediction error (MAPE) by 80% versus the prior baseline.
• Architected end-to-end MLOps pipelines on AWS SageMaker and MLflow, automating model training, versioning, and deployment across these systems.

Sharjah, United Arab Emirates

LLM adaptation research at scale — controlled fine-tuning studies, transformer architecture modification, and applied RAG for sustainability.

• Led a 22-condition LLM adaptation study comparing LoRA, QLoRA, and full fine-tuning across 4 open-weight models (Llama-3.1, Qwen2.5, Jais-2, Fanar) for Arabic medical QA, building a two-stage domain and task fine-tuning pipeline on multi-GPU AWS (4× A10G) with multi-seed reruns, bootstrap CIs, and McNemar significance testing; first-author paper under review at Cluster Computing (Springer, Q1).
• Designing controlled architecture-modification experiments on decoder LLMs (590M to 13B scaling ladder), replacing transformer FFN layers with learnable RBF/KAN variants, and built the distributed training stack (DeepSpeed ZeRO-3, 8-bit optimizers, gradient checkpointing, SLURM orchestration) enabling full fine-tuning of up to 9B-parameter models on 4× A10G GPUs.
• Built Welly, the bilingual Arabic/English RAG chatbot for MyFarmWell, a University of Sharjah–IBM app (IBM Sustainability Accelerator), answering farmer questions on irrigation, water conservation, and local regulations.
• Engineered and launched the retrieval-augmented generation pipeline over region-specific water, climate, and groundwater data, grounding LLM responses in IBM Environmental Intelligence and IBM Cloud sources for accurate, farm-level guidance.
• Co-authored 2 papers under review on ML for groundwater-level forecasting (time-series, AutoML), 2026.

Dubai, United Arab Emirates

Production LLM applications and the serverless MLOps platform underneath them, across mental health, tourism, and agriculture domains.

• Developed and integrated production-grade LLM applications for chatbots, document Q&A, and intelligent assistants using Transformers, LangChain, and RAG, achieving over 85% answer accuracy on internal mental-health domain evaluations.
• Deployed serverless AI APIs using Docker and AWS Lambda, reducing hosting costs by 25% and maintaining sub-3-second response latency under 1000 RPS.
• Built scalable MLOps pipelines on AWS SageMaker, MLflow, Airflow, and TorchServe, cutting deployment time by 30% and automating versioning, CI/CD, and monitoring.
• Fine-tuned transformer models with QLoRA and LoRA across tourism, agriculture, and healthcare, improving task accuracy by up to 18%.
• Optimized LLM summarization and multi-turn dialogue through systematic prompt engineering.

Dubai, United Arab Emirates

Classical ML and deep learning for behavioural prediction, with end-to-end ownership of the modelling workflow.

• Built and deployed machine learning models for user behavior prediction using mood analysis, psychological profiling, and sentiment analysis, improving engagement insights by 30%.
• Developed and trained classic machine learning models (random forests, XGBoost, SVMs) for structured data tasks, improving prediction accuracy by up to 22% over baseline.
• Led end-to-end ML workflows — data preprocessing, EDA, model training, tuning, and evaluation — using Scikit-learn, TensorFlow, and PyTorch, increasing model precision by 20%.
• Implemented and orchestrated experiment tracking and model versioning for better reproducibility and lifecycle management.

Sharjah, United Arab Emirates

Peer-reviewed AI research across deep learning, NLP, and clinical applications.

• Participated in and authored 7 peer-reviewed AI research papers including 2 in Q1-ranked journals with 560+ total citations and an h-index of 4, focusing on deep learning, NLP, and clinical applications.
• Developed and contributed BERT-based models and Arabic datasets for Arabic fake news detection, achieving 90%+ classification accuracy and published in high-impact NLP journals.
• Planned and evaluated deep learning architectures for Alzheimer's detection using MRI and for heart failure prediction using structured medical datasets.