Mahwiz K.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 agents, RAG systems, and fine tuned LLMs that work in production, not just in a demo. - Built a 61 megabyte text to speech model, CPU only, no GPU required, with separate male and female voices - Fine tuned and preference aligned an 8 billion parameter language model using instruction data and DPO - Trained an NER model published on Hugging Face with 95.4 percent precision and 95.5 percent recall Most AI projects stall in the same place. The demo works, then real users show up with messy data, weird edge cases, and questions the model was never trained to handle, and the whole thing breaks. I spend most of my time closing that exact gap, and the systems I build keep working after launch day, not just during the pitch. I work across generative AI, natural language processing, AI automation, computer vision, predictive analytics, and MLOps. I build production-ready AI solutions for startups, SaaS companies, healthcare, fintech, and manufacturing teams. My approach starts with the data and the constraints, not the model. Before I touch an architecture, I want to know what the input actually looks like in production, what latency and cost limits I am working with, and where the system is allowed to fail versus where it cannot. From there I build the smallest version that solves the real problem, test it against messy real world input, and only then optimize for speed and scale. Recent work includes Cadenza, a compact Urdu text to speech model that runs entirely on CPU at 61 megabytes with separate male and female voices, built through acoustic modeling, inference optimization, and conversion to ONNX. Qalb, an 8 billion parameter Urdu language model fine tuned on instruction data and aligned using DPO, with a side by side interface for comparing outputs. And a voice upgrade for a SaaS company running unattended retail machines, where I used Whisper and LLAMA2 to let customers call in problems at a vending machine and have them automatically converted into work orders. - I have shipped models most engineers will not attempt, low resource language systems where there is no existing dataset or pretrained model to lean on - I optimize for production from day one, meaning CPU inference, small model footprints, and real latency numbers, not just accuracy on a benchmark - I tell clients directly what is technically possible and what is not before any work starts, so there are no surprises halfway through a project I work with LangChain for AI agents, FastAPI for production serving, PyTorch and Hugging Face for model development and fine-tuning, and ONNX for inference optimization. I build RAG systems and custom speech pipelines using Whisper, including text-to-speech and speech recognition models. I also work across AI automation, computer vision, predictive analytics, and MLOps, with a focus on taking AI systems from prototype to production. If you're an AI startup, an enterprise team, a healthcare, fintech, or manufacturing company, or a SaaS business that needs an AI agent, AI automation, a RAG system, generative AI integration, or a model fine tuned and deployed properly, tell me what you're building and what your data looks like. I'll give you a straight answer on whether it's a fit and what it will actually take to get it running. Agentic AI, AI agents, RAG, retrieval augmented generation, LLM fine tuning, voice agents, speech recognition, text to speech, Urdu, low resource language, LangChain, FastAPI, PyTorch, Hugging Face, ONNX, Whisper, LLAMA2, DPO, preference alignment, NER, named entity recognition, prompt engineering, machine learning, production AI, CPU inference, model deployment, RAG systems, chatbot