Ahmad Sarmad Ali

PROFESSIONAL PROFILE My work focuses on building agentic LLM systems, including multi-step tool-use pipelines, autonomous workflow design, retrieval- augmented reasoning, and LLM system integration. I also bring experience across computer vision, object detection, sequence modeling, and early warning systems, enabling me to design agents that reason over both text and visual modalities. My strengths include knowledge distillation, transfer learning, transformers, machine unlearning, and federated learning/unlearning. Beyond technical depth, I bring clear communication, strong ownership, fast learning, and the ability to collaborate effectively in high-velocity environments. EDUCATION Bachelor of Science in Computer Science School of Electrical Engineering and Computer Science (SEECS) , Islamabad , 3.25 (2026) INTERNSHIP EXPERIENCE German Academic Exchange Service (DAAD) 01-Jun-2025 - 01-Sep-2025 Main Research topics: LLM fine-tuning, LLM integration, LMs as agents, Diffusion models, Generation of synthetic Docs. Using diffusion models for privacy preserving, and for Layout generation. Focused on creating a privacy-preserving synthetic Document generation framework for AI model training. Tested Qwen2.5, DeepSeek V1, and V2 as an agent to generate HTML from documents. Used Qwen2.5-32B to guide framework and diffusion models for visual elements generation and placement. Implemented LayoutDM, a discrete diffusion model for controllable layout generation of documents. WALEE 05-Aug-2024 - 31-May-2025 Main topics: Face detection, recognition, and landmark detection. Application of filters on human faces, Application of emotion-based filters. 3D asset creation from images, Face reaging filters. Implemented human face landmark detection and applied filters on faces using those landmarks in real time. Experimented with Google’s Mediapipe model for face detection and landmark detection. Contributed to the 3D-Asset generation module, developed a Fast-API for real-time 3D-asset generation. Worked on face age progression filter module using Face Reaging GANs. TUKL 03-Jun-2024 - 30-Sep-2024 Main topics: Document-AI, Machine Unlearning, Federated Unlearning, Asynchronous FL, Object Detection, Document analysis. Implemented different Federated unlearning techniques and used them for unlearning document data. Used YOLOv8 models in decentralized systems to detect different entities in documents, and then applied Federated unlearning to efficiently make the model unlearn a specific entity. Published a conference paper on Federated Unlearning for documents in ICDAR-2025 as a primary author. Presented the paper myself in CHINA; it was selected for an oral presentation. Optical Networks and Technologies Lab 05-Jun-2023 - 05-Sep-2023 Main topics: Early Warning Systems, Federated Learning, Knowledge Distillation, ML in optical Networking. Worked on Quality of Transmission Estimation Using ML in Optical Networks. Worked on QOT Estimation in Optical Networks in a decentralized system to prevent the sharing of private information of different vendors. Submitted a journal paper on federated learning based QOT estimation in the IEEE Access journal.

My work focuses on building agentic LLM systems, including multi-step tool-use pipelines, autonomous workflow design, retrieval- augmented reasoning, and LLM system integration. I also bring experience across computer vision, object detection, sequence modeling, and early warning systems, enabling me to design agents that reason over both text and visual modalities. My strengths include knowledge distillation, transfer learning, transformers, machine unlearning, and federated learning/unlearning. Beyond technical depth, I bring clear communication, strong ownership, fast learning, and the ability to collaborate effectively in high-velocity environments.