Hamza Riaz

Academic

Career

Verbatim text

Hamza Riaz
Cell: 9231589779 |  Email: hamzariaz999@gmail.com
LinkedIn: https://www.linkedin.com/in/hamzariaz970
Address: House 26-B, Street 8, Sector F-11/1 , Islamabad , Pakistan
PROFESSIONAL PROFILE
I am a final-year Computer Science student with research experience in NLP and multimodal machine learning. My primary interest
lies in Generative AI, with a focus on LLM-based systems. I have built end-to-end pipelines that combine prompt design, retrieval-
augmented generation, and structured evaluation, and I am comfortable taking a system from an initial prototype to reproducible
results. Additionally, I have certifications of Machine Learning and Deep Learning Specializations from Coursera. I have also taken all
AI-focused electives throughout my degree. This has given me a strong theoretical and practical foundation to work on SOTA
problems in this domain. Beyond model development, I can also build full stack applications to deploy and showcase these systems
outside a notebook setting. I am looking to continue this work in professional research and development environments as well.
EDUCATION
BS Computer Science
School of Electrical Engineering and Computer Science (SEECS) , Islamabad , 3.62/4.0 (2022)
INTERNSHIP EXPERIENCE
Research Intern, TUKL-NUST R&D Center (NUST x Technical University of Kaiserslautern collaboration)
03-Jun-2024 - 18-Aug-2025
- Co-developed a multimodal deep learning triage system to classify patients by the Korean Triage Acuity Scale (KTAS) using both
structured signals and unstructured clinical text. - Benchmarked multiple model families (RCNN, ACNN, BERT variants) and found
that smaller language models performed better than larger ones for this setting. - Tested LLM-based few-shot imputation to predict
missing pain scores from context, reaching performance comparable to supervised approaches. - Reframed acuity prediction as an
ordinal classification task and reported quadratic weighted kappa (QWK) alongside accuracy and macro-F1 for more meaningful
evaluation. - Built a cross-modality feature fusion module and achieved 78% validation accuracy and 0.76 macro-F1, with SHAP-
based explanations to verify clinically sensible feature importance.
FINAL YEAR PROJECT
XMedFusion: Agentic Cross-Modality Radiology Report Generation
- Building a modular, agent-based pipeline for radiology report generation that separates vision understanding, knowledge grounding,
retrieval, drafting, and refinement into clear components to output a precise and clinically accurate X-ray/CT scan report. -
Implementing structured intermediate representations, in the form of knowledge graphs, to make generation more controllable and
easier to validate than end-to-end free-form generation. - Designing retrieval-augmented workflows to pull k-most similar reports for
achieving stylistic consistency, and to reduce unsupported or inconsistent statements. - Developing robust evaluation utilities that
check output validity and measure both structure quality and end-task quality using various metrics from recent literature. - Creating a
full stack demo to showcase the system, including a fully functional backend API and a deployed web interface for interactive testing.
TECHNICAL EXPERTISE
Programming Languages
Python, C/C++, SQL. Strong fundamentals in data structures, algorithms, and writing clean, modular code.
Deep Learning Frameworks
PyTorch, TensorFlow. Model training, fine-tuning, custom modules, debugging training issues, and performance tuning.
Generative AI Tools and Techniques

AI enrichment