Muhammad Saad

Research topics: Agent Benchmarking and Agentic Operating Systems for Efficient Agent Execution on AI PC.

• Conducting research on agentic operating systems for AI PCs, focusing on how intelligent agents can coordinate tasks, tools, and user workflows in next-generation personal computing environments.
• Developing a human-agentic research assistant capable of taking a user-defined topic, decomposing it into research tasks, gathering relevant information, and generating structured, high-quality reports.
• Exploring multi-agent reasoning, task planning, tool use, and human-in-the-loop refinement to improve the reliability, usefulness, and transparency of AI-assisted research workflows.

• Deployed and customized an Open edX-based academy platform for HUMAIN Academy, including Tutor-based deployment, MySQL configuration, OAuth2 SSO integration, branded themes, enrollment workflows, and production-ready environment setup.
• Built autonomous AI agents for learning-roadmap generation, enabling users to define career or skill goals and receive structured, personalized learning paths with recommended topics, resources, and progression plans.
• Benchmarked and evaluated open-source voice cloning and text-to-speech models, focusing on inference performance, speech quality, model usability, and integration feasibility for AI-driven debate and conversational systems.
• Worked on an autonomous debate system where LLM agents generated arguments, simulated debates between multiple AI debaters, and evaluated debate quality using agent-based judging and scoring pipelines.
• Developed AI-camera research workflows for classroom intelligence, focusing on student emotion recognition, attention detection, and visual analytics to support adaptive learning and classroom engagement monitoring.

Research topics: Digital twin, Metaverse, Violence Detection, LLMs for Interactive Avatars.

• Designed and launched WudFlux, a customized virtual learning platform built on Hubs, featuring full-body avatars, real-time lip-syncing, and sentiment-aware interaction for immersive web-based learning environments.
• Built dTalk, an AI-powered avatar system with expressive 3D animations and real-time lip-sync using Mixamo and Three.js, enabling LLM-driven, speech-based interaction (GitHub).
• Built expressive 3D avatar animations using Mixamo and enabled real-time lip-sync functionality through Three.js (GitHub).
• Created a React-based analytics dashboard for the Malaria No More (MnM) project, enabling live data visualization and streamlined decision-making.
• Developed a multimodal real-time violence detection system using LSTM, GRU, and Vision Transformer architectures on Jetson Nano at the Technology Innovation Institute (TII), improving inference performance for edge deployment.

Research topics: Medical Imaging, Activity recognition, Facial emotion recognition (FER).

• Contributed to NTNU’s implementation of the facial emotional recognition module assigned by the ALAMEDA AI Toolkit to analyze facial expressions for pain assessment and emotional state monitoring in neurological healthcare.
• Attention-Based CNN-LSTM, CNN-GRU, and Video Vision Transformer (ViViT) Models for Complex Activity Recognition in Cricket.
• Teaching assistant for Python programming course.