CV
Summary of relevant experience and research
Contact Information
| Name | Muhammad Faizan |
| Professional Title | PhD Candidate (Research Assistant) |
| mfaizan@tcd.ie | |
| Location | Dublin, Ireland |
Professional Summary
My primary research interest lies in MRI reconstruction with classical and advanced deep learning models, in particular, joint MRI reconstruction and motion correction, specifically in infants.
Education
-
2025 - present Dublin, Ireland
Doctor of Philosophy (PhD)
Trinity College Dublin (TCD)
Psychology
- Thesis: Joint Reconstruction and Motion Correction of Infant Functional MRI Using Physics-Informed Deep Neural Networks
- Advisor: Prof. Rhodri Cusack
-
2022 - 2025 Islamabad, Pakistan
Master of Science (MS)
National University of Sciences and Technology (NUST)
Robotics and Intelligent Machine Engineering
- Thesis: Hybrid 3D Neural Network Architecture for Multi-Modal MRI Brain Tumor Segmentation
- Advisor: Dr. Sara Baber Sial
- Recipient of the NTF Merit Scholarship and the Prime Minister High Achievers Award (2023)
-
2016 - 2020 Peshawar, Pakistan
Bachelor of Science (BS)
University of Engineering and Technology (UET)
Mechatronics Engineering
- Thesis: Flow and Pressure Control in Variable Rate Application Sprayer
- Advisor: Dr. Muhammad Tufail Khan
- Recipient of the USAID Merit Scholarship (2017–2020)
Experience
-
2023 - 2025 Islamabad, Pakistan
Research Associate
Radar Research Lab, NUST
- Led development of RailGuard, a real-time AI safety system that fuses infrared and visible streams for track obstacle detection; introduced a novel detection-guided fusion approach that directly optimizes images for downstream detection.
- Built an open-source RRL 10k paired annotated IR–RGB dataset and achieved SOTA accuracy–efficiency trade-off (AP@0.5: 0.751) with real-time edge inference (2.5 ms Tesla T4 GPU / 3.0 ms Jetson Orin Nano, 0.04M params), reducing false negatives in adverse conditions and supporting driver-assist deployment (manuscript under review).
-
2022 - 2023 Islamabad, Pakistan
Machine Learning Engineer
Web Solutions Plus
- Trained YOLOv5 on 1M+ annotated images across five classes (person, head, car, trolley, falling person), achieving a 7.5% improvement in F1-score over the baseline; deployed the model on NVIDIA Jetson Nano for edge inference.
- Redesigned the YOLOv5 backbone to reduce redundant computation, delivering a 27% throughput improvement on NVIDIA RTX 3090 and a 3 FPS gain on Axis communication cameras.
- Integrated a centroid-based re-identification module into StrongSORT, reducing identity switches in occluded multi-person tracking scenarios and improving downstream surveillance accuracy.
-
2021 - 2021 Peshawar, Pakistan
Machine Learning Intern
National Center of Robotics and Automation (NCRA) Lab
- Trained an SSD model on a maize crop dataset, achieving an mAP@0.5 of 0.81, and deployed it on a precision agriculture sprayer for targeted pesticide application.
- Developed a custom CNN architecture to classify maize crop and weed with 97% accuracy.
Awards
-
2025 Best Paper Award
ComTech 2025
Awarded for the paper “RAG Powered LLMs for QA — Evolution, Challenges, Applications, and Future Directions.”
-
2023 Prime Minister High Achievers Award
NUST
Awarded in recognition of outstanding academic performance during MS studies at NUST.
-
2023 NTF Merit Scholarship
NUST
National Technology Fund merit scholarship awarded for academic excellence at NUST.
-
2017 USAID Merit Scholarship
UET Peshawar
Merit-based scholarship awarded by USAID for undergraduate studies at UET Peshawar (2017–2020).
-
2013 KPK Government Scholarship
GCT Attock
Provincial government scholarship awarded for academic excellence.
Publications
-
2025 RAG Powered LLMs for QA: Evolution, Challenges, Applications, and Future Directions
2025 International Conference on Communication Technologies (ComTech), IEEE
A comprehensive review of retrieval-augmented generation (RAG) for LLM-based question answering, covering evolution, building blocks, challenges, and future directions.
-
2025 3D SegUX-Net: Multi-Modal MRI Segmentation Using Large Kernels
Computerized Medical Imaging and Graphics (CMIG) — Under Review
Proposes a novel U-shaped encoder-decoder architecture using large kernel depth-wise convolution for volumetric biomedical image segmentation, achieving SOTA on BraTS 2019/2020/2023 and BTCV benchmarks.
-
2025 RailGuard: A Detection-Guided Infrared and Visible Image Fusion Framework for Enhanced Rail Safety
IEEE Transactions on Vehicular Technology — Under Review
Introduces a joint image fusion and semantic segmentation framework optimised for edge devices, achieving real-time inference on Jetson Orin Nano for railway hazard detection.
-
2022 Vehicle Recognition using Multi-Layer Perceptron and SMOTE Technique
2022 2nd International Conference of Smart Systems and Emerging Technologies (SMARTTECH), IEEE
Proposes an MLP-based vehicle recognition system using SMOTE for class imbalance, achieving 86.02% classification accuracy on the UCI vehicle dataset.
Skills
Languages
Interests
Certificates
- Deep Learning Specialization - Coursera — DeepLearning.AI
- Machine Learning - Coursera — Stanford University
- TensorFlow in Practice Specialization - Coursera — DeepLearning.AI
- Mathematics for Machine Learning Specialization - Coursera — Imperial College London
- Generative AI with Large Language Models - Coursera — DeepLearning.AI
- Introduction to ML in Production - Coursera — DeepLearning.AI
- AI for Medical Diagnosis - Coursera — DeepLearning.AI
Projects
-
3D SegUX-Net: Hybrid Neural Network for Brain Tumor Segmentation
- Designed a novel U-shaped encoder-decoder integrating large-kernel depth-wise convolution and point-wise convolution to expand receptive fields for volumetric MRI segmentation without sacrificing computational efficiency.
- Outperformed 10 state-of-the-art models on BraTS 2023: +2.18% Mean Dice over SwinUNETR and +0.29% over SegResNet, establishing a new benchmark for hybrid CNN-transformer medical segmentation.
-
Foundcog Infants Multi-band fMRI Reconstruction
- Developed a high-performance Split-Slice GRAPPA reconstruction pipeline for multiband fMRI k-space data, enabling efficient reconstruction of large-scale neuroimaging datasets comprising 139 infant subjects, over 400,000 volumes, and 3.6 million raw multiband slices.
- Established an fMRI reconstruction framework that provides clean, artifact-aware inputs for downstream motion modeling in k-space, building upon the reconstruction pipeline and forming the foundation for ongoing PhD work on joint motion correction and reconstruction to improve robustness in motion-corrupted infant and adult datasets.
-
Selected Additional Projects
- Chest Radiograph Recognition and Grad-CAM Localisation.
- Diffusion-based transformer model to denoise images with text prompt guidance.
- RAG-powered LLM chatbot for SINES NUST, integrating document retrieval with GPT-based generation.
- Sentence grammar classification with end-to-end MLOps pipeline (CI/CD, model registry, monitoring).
- Multi-layer Perceptron built from scratch using NumPy, including backpropagation and regularisation.
References
- Professor Rhodri Cusack
Thomas Mitchell Professor of Cognitive Neuroscience, Psychology. Director of TCIN, Trinity Institute of Neurosciences (TCIN). Email: cusackrh@tcd.ie