Building production-grade AI systems at the intersection of research and real-world impact. LLMs, computer vision, real-time pipelines — and the discipline of endurance sport applied to everything.
Leading end-to-end AI product delivery across multiple clients — fine-tuning LLaMA 3.1 8B for multilingual classification, engineering WhatsApp automation systems processing thousands of interactions, and deploying cinema analytics dashboards integrating live Vista SQL Server data with real-time financial reporting.
Designed a real-time OSINT incident intelligence platform for Lebanon, ingesting X, Facebook, and Telegram feeds every 2 minutes with hybrid AI + rule-based classification. Built an air quality monitoring dashboard for the French Embassy compound. Engineered a YOLOv11n fire detection system optimized for NVIDIA Jetson Nano drone footage.
Developed multimodal chatbots integrating text, voice, and image interactions using LangChain and Streamlit. Contributed to Aralects — an AI Arabic pronunciation assessment app. Researched text style transfer inspired by Meta's TextStyleBrush. Automated YouTube content pipelines for weekly podcast production.
Developed real-time object detection using YOLOv8 for a vision-based navigation system aiding visually impaired individuals. Focused on data preprocessing pipelines, model performance optimization, and evaluation methodologies for embedded deployment.
End-to-end incident monitoring pipeline for Lebanon, ingesting news from X/Nitter, Facebook, Telegram and official feeds every 2 minutes. Hybrid AI + rule-based classification (Llama via vLLM) reduced false positives in Arabic social/news data. Verified incidents were subsequently published on Al Jazeera's official Instagram account.
Full-scale AI-driven WhatsApp automation in English and Arabic. Fine-tuned LLaMA 3.1 8B using Unsloth for classification and intent recognition. Dockerized on Cloudflare domain, reducing operational costs by over 80%.
High-precision solar panel detection system using Mask R-CNN with ResNet-152 backbone, achieving 98% average segmentation confidence from aerial imagery. Delivered training sessions to candidates from Syria, Yemen, Djibouti, and Lebanon.
Real-time fire and smoke detection using YOLOv11n optimized for live inference on NVIDIA Jetson Nano. Paired with a WhatsApp forecasting bot alerting action takers when extreme fire risk is predicted within 3-day windows.
The same principles that govern elite research govern elite training — discipline, structured iteration, tolerance for discomfort, and obsessive attention to marginal gains. HYROX and functional fitness are not separate from the work. They are the same work.
Training cultivates the mental architecture that research demands: delayed gratification, sustained focus under fatigue, and the capacity to push past the point where most stop. Every session is a proof of concept.
Precision in code.
Endurance in life.
The discipline is the same.
Open to research collaborations, AI engineering roles, and ambitious product builds. If you're working on something that demands precision and performance — let's talk.
+961 78 985 105