Biography
I build computer vision and deep learning systems, with a PhD in Electrical Engineering from Chalmers and first-author papers three years running at MICCAI, the top venue in medical image computing. My work centres on structured prediction, segmentation, and efficient transformer architectures: image-to-graph models producing topologically valid vascular trees from 3D CT, and models that halve peak GPU memory at comparable accuracy. I work end to end, from data curation and training on Swedish national HPC to evaluation design and reproducible pipelines.
Education
PhD Electrical Engineering
Chalmers University of Technology (Gothenburg, Sweden)
Jan 2021 - Jun 2026
MSc. Complex Adaptive Systems
Chalmers University of Technology (Gothenburg, Sweden)
Sep 2018 - Jun 2020
BE Electrical Engineering
National University of Sciences and Technology (Islamabad, Pakistan)
Sep 2012 - Jun 2016
Interests
Experience
Doctoral Researcher, Computer Vision and Machine Learning
Chalmers University of Technology (Gothenburg, Sweden)Research on structured prediction and efficient architectures for 3D medical imaging, funded by MedTech West. Thesis: Centerline Extraction for Tubular Trees in Medical Images.
- Developed a family of recurrent transformer models for centerline graph extraction from 3D CT volumes (Trexplorer, Trexplorer Super, RefTr) producing topologically valid tree output. Three consecutive first-author MICCAI papers, the latest refining whole branch trajectories jointly to improve precision while using 2.4x fewer decoder parameters than the previous state of the art.
- Released public benchmark datasets and radius-aware evaluation metrics for centerline tracking, archived on Zenodo with a DOI.
- Built CEVAR, a fully automatic pipeline for post-EVAR follow-up CT developed with vascular surgeons at Sahlgrenska Academy, outperforming the commercial semi-automatic workflow currently in clinical use.
- Worked on efficient dense prediction: BATS (over 53% lower peak GPU memory, up to 30% faster at comparable accuracy) and ARTA (54.6 mIoU on ADE20K in the ~100M parameter class).
- Ran large-scale training on Swedish national HPC (NAISS, Berzelius, Mimer) as project PI, with sensitive patient data under ethics approval and containerised pipelines for reproducibility.
- Supervised master’s theses and taught Deep Machine Learning, Medical Imaging Systems, and Digital Health, one course in Swedish.
Research Assistant
Chalmers University of Technology (Gothenburg, Sweden)Developed an Android app for real-time on-device book detection and OCR (~250 ms) using EfficientDet and Google ML Kit, with a custom text-matching module for fast, private inventory management.Master Thesis Project
Whywaste AB (Gothenburg, Sweden)Developed an automated food inventory model using object detection and a novel object similarity network for visual clustering, leveraging synthetic data to reduce reliance on real samples and streamline training.Assistant Business Manager
Pakistan Telecommunication Company Limited (Karachi, Pakistan)Managed telecommunication infrastructure and 6,000 subscribers, leading a 24-member team to maintain and expand networks while implementing retention strategies that grew the customer base.Process Engineering Intern
Procter & Gamble (Karachi, Pakistan)Conducted PR analysis on a new packing line to enhance troubleshooting, developed a logistics tracker for cost monitoring, and streamlined processes to reduce time and improve efficiency.Sales and Design Intern
Avanceon Middle East & South Asia (Karachi, Pakistan)Applied CRM tools to improve sales at Avanceon. Gained expertise in Rockwell Automation systems and IAB software, while researching industrial applications and safety standards to strengthen technical knowledge.
Custom transformer, CNN, and recurrent architectures for structured prediction, image-to-graph modelling, segmentation, detection, and registration in 2D and 3D. Efficient and mixed-resolution designs, plus semi-supervised, self-supervised, and reinforcement learning.
Python, PyTorch, TensorFlow, CUDA, C++, MATLAB, SQL. NumPy, SciPy, pandas, scikit-learn, OpenCV, SimpleITK.
Large-scale multi-GPU training on Swedish national HPC (NAISS, Berzelius, Mimer) as project PI. SLURM, containerised environments, custom CUDA kernels, GPU profiling and memory optimisation.
Building and publicly releasing benchmark datasets for image segmentation and spatial graphs. Custom metrics, benchmark design, and failure-mode analysis. Large-scale multimodal data (CT, MRI, endoscopy, natural images, text), including sensitive patient data under ethics approval.
Git, Bash, Conda, Docker, Linux, Google Cloud. Experiment tracking with Weights & Biases and Neptune. Android development.