PhD Candidate @ Queen's Computing · BAM Lab · Kingston, Canada 🇨🇦
I work on interpretable, transferable, and explainable AI for biomedical and security-critical systems.

Portfolio · Publications · Google Scholar · LinkedIn · Email

Research badge Status badge

I build and evaluate AI systems that are not only accurate, but also inspectable, clinically meaningful, and robust under messy real-world conditions.

My current research lives at the intersection of:

• 🩻 Biomedical Vision-Language Models — evaluating and improving medical VLMs for radiology.
• 🔍 Explainable AI — using tools like Grad-CAM, SHAP, probing, and error analysis to understand model behavior.
• ⚖️ Long-tailed and imbalanced learning — especially in healthcare, where rare findings matter.
• 🧬 NLP for biomedical evidence — literature screening, clinical text classification, and EHR-style data analysis.
• 🛡️ Biometrics & cybersecurity — face presentation-attack detection, liveness detection, and adversarial robustness.
• 🧭 Research tooling — building interfaces that help researchers navigate papers, models, experiments, and knowledge graphs.

Medical AI should not be a black box that only returns a label.
It should explain what it saw, why it matters, and where it may fail.

Right now, I am especially interested in:

• 🩻 Interpretable biomedical VLMs for chest radiology
• 📉 Failure modes of zero-shot medical AI under dataset imbalance
• 🧠 Knowledge-guided and graph-aware medical AI systems
• 🧪 Evaluation protocols for robustness, grounding, and clinical usefulness
• 🗂️ Tools for organizing research knowledge and project memory

A few recent pieces of work I am excited about:

• DepthPulse+: A Depth and Vital Sign Based Method for Face Presentation Attack Detection IEEE ICC 2026 Depth maps + remote photoplethysmography for robust liveness detection against print, replay, and mask attacks.

• Interpreting Biomedical VLMs on High-Imbalance Out-of-Distributions: An Insight into BiomedCLIP on Radiology BioKDD 2025 Investigates how biomedical VLMs behave under imbalanced, out-of-distribution chest X-ray settings.

• Visualizing Temporal Interval Hierarchies NASA Formal Methods 2025 A visualization system for streaming PID-controller data and temporal interval hierarchy analysis.

• Vulnerability of Open-source Face Recognition Systems to Blackbox Attacks: A Case Study with InsightFace IEEE CICS 2023 Studies black-box adversarial attacks and transferability in open-source face recognition systems.

📖 Full list: Google Scholar · Publications page

research_stack = {
 "medical_ai": ["BiomedCLIP", "BioViL-T", "VLM evaluation", "radiology"],
 "xai": ["Grad-CAM", "SHAP", "attention analysis", "failure-mode analysis"],
 "ml": ["PyTorch", "TensorFlow", "scikit-learn", "long-tailed learning"],
 "nlp": ["BERT", "clinical text", "literature screening", "LLM-as-judge"],
 "tools": ["FastAPI", "React", "D3", "SQLite", "Socket.IO"]
}

I like projects that combine:

1. A real problem — preferably messy, imbalanced, and clinically or socially meaningful.
2. A strong technical question — not just "can we get higher accuracy?"
3. A careful evaluation plan — including uncertainty, bias, and failure cases.
4. Readable outputs — figures, tools, dashboards, explanations, and documentation.

Nafiz's GitHub stats Top languages

Outside of models and manuscripts, I enjoy building tools that make research less chaotic, turning rough ideas into structured systems, and explaining complicated papers in a way that actually makes sense.

I also have a soft spot for projects that start with:

"This is probably a bad idea, but what if we tried it?"

Those are often the fun ones. 😄

Always happy to connect about medical AI, XAI, VLMs, research tools, and PhD-life debugging.
sadman.n@queensu.ca