Almaan Khan

About

I'm a Computer Science student at King Fahd University of Petroleum and Minerals (KFUPM), class of 2027, with a 3.961 / 4.0 GPA. I'm a Research Fellow with the Fatima Fellowship, working on unsupervised 3D brain MRI anomaly detection under Dr. Sarah Anjum (Santa Clara University). I previously interned at BestCircle AI and the Saudi Medical AI Lab (SMAIL).

My research spans medical vision-language model safety, generative models for medical imaging, interpretable dyslexia screening, and rare-object detection. My work includes an accepted paper at the NeurIPS GenAI4Health Workshop (2026) and a dyslexia detection paper in the Journal of Undergraduate Research International (JURI). I was selected among the top 100 students from over 16,000 applicants at KAUST Academy's AI Summer School, and received the University of Toronto's $100,000 International Scholar Award.

I also build applied AI systems: Arabic feedback analysis for the National Center for Mental Health Promotion, document-grounded compliance workflows, and agentic shopping assistants. I'm the Software Team Lead for KFUPM's University Rover Challenge team and a member of its VEX U Robotics Team.

// stack

Python PyTorch TensorFlow MONAI LangChain LangGraph OpenCV Pandas NumPy SQL PostgreSQL Vertex AI Next.js FastAPI Git

Publications

2026
NeurIPS GenAI4Health
Accepted workshop paper

Rejected When Asked, Followed When Assumed: False Presuppositions in Medical Vision-Language Models

Sultan Alshehri, Almaan Khan, Sarah Alharbi, and Marios Savvides

Medical VLMs often describe findings they previously judged absent when a new question assumes those findings are present. Across five open models, this occurs in 82.9–100% of such cases; a frontier reasoning system does so in 72.2%. Verification prompts help only partly. Fine-tuning can reduce the error while also shifting direct presence judgments, so lower error rates alone do not establish better visual grounding.

read on OpenReview read paper (PDF)

09 / 2026
JURI · 2(3A), 70–76

Feature-Efficient and Interpretable Dyslexia Detection via Soft Voting Ensemble Learning

Almaan Khan and Shujaat Khan

Combines AdaBoost, Gradient Boosting, XGBoost, and CatBoost with mRMR feature selection and SHAP explanations. Using 165 features and stratified 10-fold cross-validation, the ensemble achieved 91.6% accuracy, 90.7% precision, 91.6% recall, and a 90.9% F1 score. Published September 27, 2026 in the Journal of Undergraduate Research International.

read journal article read paper (PDF)

Research & Experience

04 / 2026 — present

Research Fellow, Fatima Fellowship

Investigating unsupervised 3D brain MRI anomaly detection under Dr. Sarah Anjum (Santa Clara University) using MONAI generative models. Developing hierarchical autoencoder feature residuals, with current tumour-localisation AUROC improving from 0.556 to 0.648; benchmarking L1, L2, cosine, and Mahalanobis scoring.

[generative models]

06 / 2026 — 08 / 2026

AI Engineering Intern, BestCircle AI

Co-developed a PDPL compliance platform with Next.js, PostgreSQL, and Vertex AI, covering document ingestion, evidence retrieval, citation-grounded findings, compliance classification, and human review. Also evaluated joint-source dependence in five VLMs across 2,532 questions, finding 0.6–1.9% confirmed dependence.

[ai engineering]

09 / 2025 — 08 / 2026

Research Intern, Saudi Medical AI Lab (SMAIL)

Worked with Dr. Sultan Alshehri (Carnegie Mellon University) on medical VLM hallucinations, developing a LoRA-tuned verification gate to reject false-premise queries. Related work is described in the accepted NeurIPS GenAI4Health Workshop paper above.

[medical ai]

01 / 2025 — 12 / 2025

Undergraduate Researcher, BRAIN Lab

Investigated class imbalance and rare-object detection under Dr. Muzammil Behzad. Benchmarked five modern detector architectures on MS-COCO, comparing class-specific confidence thresholds and ensemble techniques across rare and frequent classes.

[computer vision]

01 / 2025 — 06 / 2025

Undergraduate Researcher, SDAIA-JRCAI

Worked with Dr. Shujaat Khan on dyslexia detection, reproducing the PLOS ONE gamified-test baseline before developing a soft-voting ensemble with mRMR feature selection. This work became the JURI publication above.

[interpretable ml]

06 / 2025 — 09 / 2025

Top 100 / 16,000+, KAUST Academy AI Summer School

Selected from a national applicant pool to build agentic AI systems and applied ML projects under intensive mentorship from KAUST faculty and industry researchers.

[selection]

Leadership & Robotics

08 / 2026 — present

Software Team Lead, KFUPM University Rover Challenge Team

Leading software development and autonomous algorithms for KFUPM's competition rover.

[autonomous systems]

08 / 2026 — present

Member, KFUPM Robotics Team (VEX U)

Selected to represent KFUPM in the VEX U Robotics Competition.

[robotics]

01 / 2025 — 01 / 2026

Marketing Team Lead, IEEE KFUPM Student Branch

Led five members in planning and running campaigns for IEEE events attended by 300+ students over two semesters.

[leadership]

Awards & Recognition

2025

Dean's List & Calculus Award, KFUPM

Recognised among the top 10% by CGPA in the College of Computing and Mathematics (terms 241, 242, and 251), and awarded for academic excellence in MATH201 (Calculus III).

[academic honours]

2024

Physics Excellence Awards, KFUPM

Awarded by the University President for academic excellence in PHYS101 and PHYS102 (General Physics I and II).