At Qubeterra, I designed an extensible multi-agent architecture with reusable agents, personas, and LLM reasoning workflows. I also implemented multi-provider LLM integration and exposed agent capabilities through FastAPI APIs.
For my Decision Discussion Platform project, I developed a multi-agent system where specialized AI personas research evidence, debate, and produce evidence-backed recommendations. Its RAG pipeline uses BGE-M3 embeddings, a cross-encoder reranker, and Supabase pgvector.
For BASEER, my graduation project, I owned the fabric-inspection module and trained an STPM anomaly-detection model on about 5,000 images, achieving 96.3% AUROC. I deployed real-time defect scoring and heatmap generation through FastAPI microservices, using ONNX Runtime for inference.
At the National Telecommunication Institute (NTI), I built and optimized real-time YOLOv5 detection and tracking pipelines for football match analysis. I also built a machine-failure prediction API using XGBoost, scikit-learn preprocessing, and FastAPI.
