Ali Subhan

At FRI, University of Ljubljana, I deployed ScholarAgent, an enterprise-grade multimodal AI assistant for scientific Q&A. I fine-tuned Qwen3.5-9B with QLoRA, served it via vLLM, and designed a Supervisor/Subagent architecture coordinating arXiv, Semantic Scholar, and live PDF-ingest agents.

I built a two-stage RAG pipeline using Qdrant HNSW and cross-encoder reranking, improving top-3 retrieval precision by 12%. I also implemented LLM-as-a-Judge evaluation for response quality and citation-faithfulness monitoring across the agentic pipeline.

Previously at Proto IT Consultants, I built Python agentic automation pipelines with LangGraph and MCP, connecting OpenAI and Anthropic APIs with CRMs, Zoom, QuickBooks, Stripe, and Google Workspace. My work included meeting and calendar assistants, resilient workflow monitoring, and a multi-agent social-media content pipeline.

At NUST's OPTIMAL Lab, I developed GaitID, a real-time gait-recognition access system reaching 95% accuracy at 28 FPS on Jetson Nano. I also built an on-device Edge AI inspection pipeline using YOLOv8 and TensorRT that delivered 3x faster inference.