Muhammad Fazeel

Grenoble INP Phelma, Université Grenoble Alpes, Grenoble, France

• Master's thesis: Automatic Quantification of Skeletal Muscle in Pediatric MRI, carried out at GIPSA-Lab (CNRS) with Inria Grenoble Rhone-Alpes and CHU Grenoble Alpes. Degree awarded 2026, mention Assez Bien.
• Key modules: Machine Learning, Image Analysis and Computer Vision, Image Processing, Signals and Systems, Bio and Neuro Imaging Methods, Scientific Programming.

National University of Sciences and Technology, Islamabad, Pakistan

• Key modules: Artificial Intelligence, Computer Vision, Machine Learning, Mobile Robotics, Robot Mechanics and Control.

Turing, remote

• Contracted to provide autonomous, remote data science services for a global technology client, supporting a flagship large language model product with a user base above 500 million.
• Optimise the model's agentic capabilities in Python, and previously processed, evaluated and structured large-scale datasets to establish performance metrics for the system.

GIPSA-Lab (CNRS), Inria Grenoble Rhone-Alpes and CHU Grenoble Alpes, Grenoble, France

• Quantified how far SegmentAnyMuscle, a muscle-segmentation foundation model pretrained on adult MRI, transfers to pediatric MRI, and characterised its failure modes along acquisition, anatomical coverage and pathology on a clinical cohort of nine children.
• Benchmarked three annotation-efficient adaptation routes: parameter-efficient fine-tuning of the foundation model with a frozen Vision Transformer backbone, a trainable mask decoder, adapters and a mixture-of-experts head; a 2-D nnU-Net trained from scratch, which segmented a held-out severe-pathology thigh at Dice 0.94; and self-supervised masked-image pretraining on 123 unlabelled MRI series.
• Tested MedGemma, a medical vision-language model, as a controller for pixel-level annotation in two agentic settings, tool-driven control of brush and eraser actions and function-calling-driven segmentation, to establish where a general medical VLM stops being reliable for localisation.
• Built a human-in-the-loop annotation workflow in which clinicians corrected model predictions instead of redrawing them, producing 285 expert-reviewed labelled slices across 30 volumes from a previously unannotated clinical archive.
• Designed the evaluation protocol for the degenerate case of an empty reference mask, replacing Dice with the muscle-to-total-tissue ratio and explicit false-positive accounting, delivered as a reproducible quantification module. Python, PyTorch, nnU-Net v2, NVIDIA H100 on the GRICAD HPC cluster.

Khyber Medical University, Peshawar, Pakistan

• Developed deep learning methods on gigapixel whole-slide histopathology images to reduce subjectivity in cancer grading, under a research grant co-funded by the Higher Education Commission Pakistan and the British Council UK.
• Ran GPU-accelerated sequencing pipelines with NVIDIA Clara Parabricks, aligning, calling and annotating variants from raw FASTQ data, with downstream analysis to prioritise candidate oral-cancer genes.

AI in Healthcare Lab, National Center of Artificial Intelligence, Peshawar, Pakistan

• Built deep learning models for atrial fibrillation, bradycardia and tachycardia classification from ECG signals.
• Collaborated on sound source separation and audio denoising over spectrograms using generative adversarial networks.

SpectraWeaver, an agentic image-processing assistant

Streamlit application built on Gemini 2.5 Pro that turns a stated preprocessing objective into executable Python: it inspects each image, writes and runs the processing code in a constrained local environment, visualises the result and evaluates whether the objective was met. Multi-turn conversations branch from any earlier message, so alternative pipelines can be compared without losing history. github.com/fazeel15/SpectraWeaver

Interpretable Broders' grading of oral cancer

Detection and classification of squamous cells in histopathology images with object-detection networks, producing case-level grades supported by pathologist-interpretable evidence rather than a single opaque score.

Arrhythmia classification on PhysioNet/CinC 2020

ECG preprocessing, handling of class imbalance, and evaluation of multi-class cardiac arrhythmia models with metrics appropriate to the imbalance.

Sequencing pipeline for oral-cancer gene discovery

Alignment, variant calling and annotation of Illumina next-generation sequencing data with NVIDIA Clara Parabricks, then downstream analysis and visualisation to prioritise candidate genes.