Muhammad Waleed B.

I'm happy to start with a free consultation, quick POC, or a test task, your call. See the quality first, then decide. I build production AI systems: computer vision pipelines, RAG knowledge bases, LLM fine-tuning, voice and chat agents that run at scale for market giants serving millions of customers. $300K+ earned across 80+ Upwork contracts and 4500+ hours, 100% Job Success, Top Rated Plus. I lead the AI engineering team at AB Ark. WHAT I BUILD Computer vision Object detection and tracking (YOLO, OpenCV), CCTV and video analytics, edge inference on NVIDIA Jetson, facial expression and body-language models, OCR and document layout analysis, image segmentation. RAG and knowledge systems Private document brains over Google Drive, SharePoint and internal wikis, with citations back to source. Vector search (pgvector, Pinecone, Qdrant), hybrid retrieval, re-ranking, multi-LLM routing, document classification and extraction. LLM engineering Fine-tuning and LoRA training, prompt architecture, evaluation harnesses so you can measure whether a change helped, structured output and schema enforcement, GPT, Claude and open-weight model integration. Voice and conversational AI Real-time voice agents on Twilio, Telnyx, Retell and LiveKit including human-like interruption handling and warm transfer to a live agent. AI agents and automation LangChain and LangGraph agents with tool calling, multi-step workflows, retrieval and human-in-the-loop approval steps. Deployment and MLOps Docker, Kubernetes, CI/CD, AWS and GCP, model serving, monitoring and drift detection. FastAPI and Django when the model needs an API around it. RECENT WORK Edge video analytics on NVIDIA Jetson: real-time object detection on CCTV streams for an on-premise deployment Private RAG "Knowledge Brain" over Google Drive with SOP indexing and citation-backed answers Computer vision SaaS for CCTV footage analysis, built as a multi-tenant product Telnyx voice assistant with warm transfer and human-like interruption handling LLM/RAG document classification and workflow design Computer vision models for facial expression, body-language analysis or custom object detection STACK Python · PyTorch · TensorFlow · OpenCV · YOLO · Hugging Face Transformers · spaCy · scikit-learn · LangChain · LangGraph · LlamaIndex · OpenAI · Anthropic · pgvector · Pinecone · FastAPI · Django · PostgreSQL · Docker · Kubernetes · AWS · GCP · NVIDIA Jetson HOW I WORK Discovery first. I define the data, the model approach and the evaluation metric before writing training code, so "done" is measurable rather than argued about. Milestones with written acceptance criteria, or hourly with daily updates. Your choice. You own the code, the model weights and the infrastructure. NDA and IP assignment on request. Send me your dataset, your accuracy target, or the pipeline you have now, and I will come back with an approach, the risks, and an estimate.