Elia Ghazal

SHARP-RAG: Self-Correcting Hierarchical Agentic Retrieval-Augmented Generation for Multi-Hop Question Answering

Elia Alghazal · Independent Researcher · Beirut, Lebanon

Multi-hop question answering requires chaining evidence across several documents, a setting in which naive RAG frequently fails because it retrieves once and never verifies whether the retrieved context supports an answer. SHARP-RAG addresses this with a four-agent LangGraph pipeline: a Planner, Retriever, Critic, and Synthesizer cooperate in a cyclic stateful graph where the Critic emits a structured JSON verdict that gates answer generation and drives targeted re-retrieval. Evaluated on 20 HotpotQA fullwiki questions, SHARP-RAG underperforms both baselines.

Finding: the self-correcting loop underperformed naive RAG, driven by over-triggering from the 8B evaluation critic standing in for the intended 70B model, not the architecture itself. Numbers below.

Dialect Is Not Error: Auditing Prestige Bias in LLM-as-a-Judge Systems Under Arabic Diglossia

Elia Ghazal · with William Ishak · advised by Dr. Charbel Boustany, AUST

A benchmark-driven audit of whether LLM-as-judge systems evaluate Arabic text on merit or on register: does a response written in a regional Arabic variety get scored differently than an equally correct response written in Modern Standard Arabic. Multiple LLM judges audited across pointwise and pairwise comparisons, with a native-speaker validation pass on the dataset itself. Full methodology and results to follow once the paper clears review.

CrashLens: Smart Crash Detection and Emergency Response via IoT and Artificial Intelligence

E. Alghazal, G. Khayat, W. Ishak, B. Farhat, M. Allaw · Advised by Dr. C. Boustany, AUST

CrashLens is an embedded crash-detection and emergency-dispatch platform. A Raspberry Pi 5 with a six-axis IMU polls acceleration at 100 Hz and confirms a crash by sustained threshold, not vision inference, then dispatches an SOS over LTE on telemetry alone, before any video has uploaded, cutting measured dispatch latency from roughly 14 minutes to about 30 seconds in bench tests. Video and key frames upload afterward on a separate path, triggering a forensic stage that runs Gemini multimodal video analysis in parallel with license-plate OCR, cross-checked against Lebanon's vehicle registry, surfaced through a Flutter driver/responder app and a Next.js insurance dashboard.