Posts by Collection

Agentic Process Discovery

Reverse-engineers hidden administrative workflows from email and chat history using a two-stage local LLM pipeline with no data leaving the device.

Membership Inference Attacks on Quantized LLMs

NYU Abu Dhabi capstone testing whether post-training quantization protects LLMs against membership inference attacks, evaluated on Pythia-12B across six PILE domains. Currently under peer review at a security-focused academic venue.

Gmail RAG Pipeline

Local semantic search over Gmail using whole-email BGE embeddings and ChromaDB, with no email content leaving the device. Benchmarked against a BM25 baseline. Foundation for the Agentic Process Discovery system.

AiXplore

Multi-agent MENA travel planning assistant built at a hackathon on the aiXplain platform. Five-agent pipeline: input parsing, restaurant search, attraction search, itinerary generation, and validation with Arabic output.

MS Lesion Evaluation Framework

Lesion-centric evaluation framework (MSEval) that exposes failure modes hidden by standard Dice metrics in MS lesion segmentation models. Published at IEEE IJCNN @ WCCI 2026.

Python Generated Animation

Automatically and Efficiently Generating Animations to Aid Teaching

Mixed-Precision Quantization

Mixed-precision quantization framework for DNNs and LLMs using RL-based per-layer bit-width search, extending ANT and OliVe. Research work at eBrain Lab, NYU Abu Dhabi.

Rethinking Evaluation of Multiple Sclerosis (MS) Lesion Segmentation Models

Published in IJCNN SS46 Computationally Intelligent Techniques in Early Prediction and Detection of Brain Disorders, 2026

Note: This paper has been officially accepted for publication at IJCNN 2026 (SS46: Computationally Intelligent Techniques in Early Prediction and Detection of Brain Disorders). The final version has not yet been formally published. We will update this page with official links, slides, and the final manuscript once it is published.