Building production-grade AI systems at the intersection of research and real-world impact. LLMs, computer vision, and real-time pipelines designed for meaningful outcomes.
Working remotely as an AI Developer with Nucleus Ventures, based in Nicosia, Cyprus.
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.
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