Muhammad Abubakar Saddique

Designed end-to-end AI and MLOps architectures for enterprise products, delivering AI solutions including a Real-time Avatar Virtual Assistant, AI Token Factory, and computer vision security platforms. Optimized and productionized AI models through inference tuning, deployment automation, and scalable MLOps practices.

Architected 13 continuous learning pipelines for LLMs, VLMs, YOLO, and traditional ML models via Kubeflow, replacing fragmented legacy workflows with a fully automated retraining infrastructure. Migrated legacy Kubernetes deployments to a KServe multi-model serving platform with Knative, KEDA, and KAI scheduler, improving throughput and GPU utilization by 40%.

Architected and deployed a Retrieval-Augmented Generation (RAG) system to improve knowledge retrieval precision and enhance AI-powered e-learning experiences. Engineered agentic AI pipelines for automated student query resolution, reducing manual support overhead and improving response efficiency.

Built a 10K-image domain-specific dataset of Pakistani visual elements through large-scale Instagram scraping and curation for generative model training. Fine-tuned Stable Diffusion 1.5 and LoRA adapters, integrating with LLaMA 3.1 to improve prompt alignment and output quality.

Implemented the Stable Diffusion architecture from scratch, designing U-Net, VAE, and DDPM modules to study the mechanics of image generation from noise. Reconstructed the mathematical foundations of diffusion models, analyzing noise scheduling, probabilistic sampling, and the role of Linear Algebra, Calculus, and Probability in generative processes.

Conducted a one-month Python training program for 30+ students, covering fundamentals up to pre-OOP concepts. Guided learners through two hands-on projects while emphasizing problem solving, coding best practices, and project management.

Grade: 3.69

Bachelor of Science in Artificial Intelligence with a CGPA of 3.69, graduating as a Gold Medalist.