Posts by Collection

portfolio

flwr-trees: Federated Learning for Tree Ensembles

An open-source Python library (on PyPI) that trains Random Forest, XGBoost and gradient-boosted tree models across clients using the Flower framework, behind a standard scikit-learn interface. Introduces histogram-based aggregation that cuts the first-round upload by 45–83% on benchmark datasets.

Contactless Vital-Sign Estimation from Video (rPPG)

Estimating heart rate, heart-rate variability and SpO2 from an ordinary webcam using remote photoplethysmography, as part of the “Walk for Life” project at the Institute of Liver and Biliary Sciences, New Delhi (2026–present). Public prototypes: SerenePulse and an in-browser HRV dashboard.

Brain-to-Brain Decoder: Dyadic EEG Representation Learning with CEBRA

An independent study that learns joint neural embeddings of two people in conversation from hyperscanning EEG. A hybrid CEBRA embedding decodes affective condition with 0.967 k-NN accuracy, while label- and temporal-shuffle controls fall to chance. Includes a reproducible pipeline and a plan for scaling to multiple dyads.

LUFY: Retrieval-Augmented Understanding of Legal Documents

A legal-document assistant that turns an uploaded contract into a plain-language summary, a three-tier clause risk analysis and citation-grounded Q&A in English and 16 Indian languages. Uses hybrid dense + BM25 retrieval with reciprocal rank fusion and structure-aware chunking.

publications

Low-Resource Indian Sign Language Video Generation via Compressed Multi-Condition Tokenisation

B.Tech Thesis, Department of Computer Engineering, Jamia Millia Islamia, 2026

B.Tech thesis (with Hammad Ali). The first text-to-skeleton-to-video baseline for Indian Sign Language on ISL-CSLTR (471 videos), using body-part finite scalar quantisation and a frozen Stable Diffusion + ControlNet + IP-Adapter renderer.

Recommended citation: H. Ali and M. A. Khan. (2026). "Low-Resource Indian Sign Language Video Generation via Compressed Multi-Condition Tokenisation." B.Tech Thesis, Jamia Millia Islamia, New Delhi. Supervisor: Mr. Shamim Ahmad.
Download Paper

Positional Bias in Nested Defeasible Reasoning: Order Sensitivity and Unfaithful Intermediate Steps in Small Language Models

Manuscript in preparation, 2026

Do language models reason over nested defeasible premises in an order-invariant way, and do their correct answers rest on correct intermediate steps?

Recommended citation: M. A. Khan. (2026). "Positional Bias in Nested Defeasible Reasoning: Order Sensitivity and Unfaithful Intermediate Steps in Small Language Models." Manuscript in preparation.

FIB-4 False Negatives in MASLD: Clinical Phenotypes and an Augmented Triage Model

APASL Single Topic Conference (STC) 2026, Kumamoto, Japan. Oral presentation FP5-2, 2026

Among MASLD patients classified as low-risk by FIB-4 (NHANES 2017–2020), 16.6% had significant (F2+) fibrosis, rising to 37.7% at BMI ≥ 40 kg/m². A supplementary model, FIB-4 Plus, outperformed the NAFLD Fibrosis Score (AUROC 0.837 vs. 0.749, DeLong p = 0.019). Oral presentation; APASL International Travel Award.

Recommended citation: M. A. Khan. (2026). "FIB-4 False Negatives in MASLD: Clinical Phenotypes and an Augmented Triage Model." Oral presentation FP5-2 (Abstract No. 10340), APASL Single Topic Conference 2026, Kumamoto, Japan.
Download Slides

talks

teaching

Technical Workshop Series, IEEE JMI

Organiser (General Secretary), IEEE Student Branch, Jamia Millia Islamia, 2024

General Secretary of the IEEE student branch at Jamia Millia Islamia (March 2024 – present; earlier a member of the graphic design team, March 2023 – February 2024). As General Secretary, I led a team of 15+ student volunteers that organised 10+ technical workshops, reaching more than 1,200 students in total.

Introduction to Machine Learning Workshop

Workshop, Co-instructor, W3B (AI and Blockchain Society), Jamia Millia Islamia, 2025

30 January – 8 February 2025. One of a team of instructors who taught an introductory machine-learning workshop to 150+ undergraduate students.