Hafiz Umar I.Private profile: Upwork shows its earnings only to signed-in members, so we check it every few weeks, not daily. MRR here is a rough estimate (Upwork's six-month earnings ÷ months active). Last read Nov 5, 2026.

I build AI systems that run in production: LLM agents that read and act on real business documents, and computer vision that watches live video around the clock. 10+ years in ML and AI, an MS from NUST and a paper in IEEE Access. AGENTIC AI AND LLM SYSTEMS Multi agent workflows with tool calling, routing, verification steps and human approval where it matters. RAG over large document sets with vector search (Milvus), backed by evaluation sets so quality is measured, not guessed. Fine tuning open models with LoRA and QLoRA, serving with vLLM, TGI or Triton. I work with Claude, GPT, Gemini, Llama and Mistral, mostly in Python with LangChain and FastAPI. COMPUTER VISION AND VIDEO ANALYTICS Detection, tracking and segmentation with YOLO, SAM and Detectron2. Multi camera event detection over RTSP and Kafka: PPE, fire and smoke, violence, intrusion and face tracking. Industrial inspection, gauge reading, OCR and document AI, plus vision language models such as LLaVA for richer image understanding. PRODUCTION ENGINEERING FastAPI microservices, Docker, AWS (EC2, SageMaker, Lambda, S3), multi GPU training and inference, and monitoring of accuracy, cost and latency after launch. RECENT WORK • Lead a small engineering team building a multi agent document intelligence platform in production, with LLM verification of every extraction and model routing to keep inference costs down. • Real time safety monitoring on 12 to 16 live camera streams per vessel, deployed and running at sea. • OCR and LLM pipeline over 10,000+ engineering manuals, with fine tuned Vicuna and T5 models. • Trained and served vision models on 8M+ images across a 13 server GPU setup. HOW I WORK I start by pinning down the problem and what "good" means for you, then get a working version in front of you quickly so we can judge something real. You get clean code, tests where they matter, an evaluation you can rerun, and documentation the next engineer can follow. Clear updates, honest estimates, no surprises. PUBLICATIONS • Deep Learning Based Correct Answer Prediction for Developer Forums, IEEE Access, 2021 • Machine Learning Based Classification of Septic Shock Using Microarray Gene Expression Data, FIT 2019 If you have a problem where AI could save real hours or catch what people miss, send me a message and tell me about it.