Umer S.

Across 12 years of software development and 9 years specializing in machine learning, I have built systems ranging from conversational chatbots to computer vision based grading tools, working with clients across different industries and problem types. What stays consistent across all of it is the same goal, moving past a model that performs well in isolation to one that performs well once it is actually in production. I have completed 104 jobs on Upwork, earned $200K plus, and maintained a 100% Job Success score with Top Rated status. Clients have used my work to ship products up to 3x faster while cutting $10K a month in engineering costs, using AI assisted development on top of solid computer science and machine learning fundamentals. I also hold a Master's degree in Machine Learning, and I currently lead a 45 engineer team building production AI and ML systems. What I actually build: Conversational AI and LLM systems. Chatbots and multi turn assistants built on RAG pipelines and vector search, with real guardrails so the model does not make things up when it does not know the answer. Computer vision pipelines. Detection, classification, and grading systems trained on real world data, not clean benchmark datasets. Natural language processing. Text classification, sentiment analysis, entity extraction, and low resource language OCR, including work on Urdu, where off the shelf tools mostly do not exist. Production ML infrastructure. Models packaged, monitored, and deployed with Docker, Kubernetes, and cloud infrastructure on AWS, Google Cloud, or Azure, including edge deployment where latency or hardware is constrained. Some of what my team has shipped: ✦ Avatare, a conversational AI platform handling multi turn dialogue in production ✦ Compugrade, an AI assisted autograding system for K12 education built on computer vision ✦ AffectCX, a live conversational chatbot supporting pharma sales teams ✦ A multimodal conversational AI platform combining RAG based retrieval with generation ✦ A computer vision grading pipeline for physical collectibles, handling classification and defect detection ✦ Urdu OCR, a low resource language OCR system built where almost no existing tooling was available Technical stack: Python, PyTorch, TensorFlow, Keras, JAX/Flax, OpenCV, Docker, Kubernetes, AWS, Google Cloud, Azure, ONNX, TFLite, CoreML AI and ML methods I work with: Large language models including GPT4, Claude, and Llama 2. Retrieval Augmented Generation and vector databases. Traditional ML including random forests, SVM, and KNN. Deep learning including CNNs, RNNs, GANs, transformers, and BERT. MLOps including CI/CD, model monitoring, and versioning. If you are dealing with a model that only works some of the time, a system that needs to move from prototype to something production ready, or a computer vision or NLP problem that off the shelf tools have not solved, message me. I typically reply within 2 hours and can prioritize urgent projects. I will tell you honestly whether AI is the right tool for the job before I agree to build anything.