List of Developed Prototypes

List of Developed Prototypes

A Red Team Prototype for Harmful Content Detection

This demo showcases a system that verifies the faithfulness of a summary to its original text. As part of a red-teaming project, it uses multiple LLMs as judges to assess reliability. The interface includes fields for text and summary, plus a button to run the verification

‎

 

Arabic Text Style

An academic tool designed for common tasks like text summarization, QA, translation, code generation, and retrieval. Its main feature is that instructors can control and manage its use, selecting specific references or language models, unlike general tools that lack classroom control.

‎

 

Instruction Tuning for LLM and the Role of Machine Translation

Tawjeeh is a platform for designing, drafting, reviewing, and publishing prompts for instruction tuning experiments. The platform curated datasets from Huggingface platform. Based on this, it selects a sample of the dataset where the prompt designer will see how his prompt is being rendered for a given dataset sample. Once he is satisfied with his prompt, he will submit it for review. A review workflow will be then applied to his submitted prompt where the prompt can undergo the following statuses: draft, submitted, returned for modification, or approved. Once the prompt is approved, it can be accessed for read-only. Approved prompts can be shared as a dataset for further experimentation.

 

Model Generated Text Detection at Scale

With the current surge of AI models that generate text with near-human performance, distinguishing whether a text is generated by AI or written by humans becomes a task with high impact. In this project, we developed deep-learning-based methods to classify human-written text from AI-generated text. For generalization, we incorporated AI text generated from multiple AI models. Our experimentation spans multiple domains with various prompting methods. Within our experimentation settings, our trained detectors achieved high performance in detecting whether the text is AI-generated or not.


Tafsir: Harnessing LLMs for Concise and Relevant Tafsir Extraction

This prototype aims to provide concise summaries, answer queries, offer deep insights, deliver contextual details, perform comparative analyses, support vocabulary enrichment, and enable interactive exercises to enhance user engagement and comprehension.

 

AI-Based PV Panels Inspection using Drones

This video demonstrates an AI-powered drone-based system for inspecting photovoltaic (PV) panels. It highlights both real-time and offline inspection features, including a live video stream dashboard for immediate monitoring and an offline workflow for uploading images and videos. The system processes thermal images to detect issues like hotspots and anomalies, generating detailed reports with historical data for trend analysis and maintenance. It also includes secure user management with authentication and authorization, plus customization options to fit specific user needs.

 

Computer-Aided Diagnostic System for Skin Lesions

This demo video shows the basic functionality of the first design iteration (v0.1.0) for the skin lesion classification application from the use-case project (Computer-Aided Diagnostic Systems for Skin Lesions)

 

Continuous Saudi Sign Language Recognition

A platform to recognize the Saudi Sign Language aiming to improve the accessibility and communication of the deaf community.

 

 

Showcase/MVP of Personal Protective Equipment (PPE)

Video showing a live feed from one of the factory sites, monitoring workers’ compliance with safety protocols. The system also displays compliance data over time, provides alerts to management, and reports any violations.

 

 

Face Recognition

A face recognition system for event attendance can be deployed on a compact NVIDIA Jetson Orin device. The camera detects faces in real time, and the system extracts facial embeddings for comparison against a vector database of registered participants. Once matched, the attendee’s name appears on a display alongside their detected face, confirming attendance instantly. This setup provides a smooth and automated check-in process, reducing manual effort while enhancing accuracy and efficiency for event organizers and participants alike.

 

 

 

Multimodal Deep Learning for People Detection and Demographic Analytics

This prototype represents a multimodal deep learning system to detect key demographic attributes of speakers, including dialects, gender, and age groups. By leveraging multiple modalities, such as text, speech, and video, the system aims to provide comprehensive demographic analytics in real-time

 

 

 

Academic Tool

A versatile academic assistance platform offers a range of essential tools, including code generation, translation, entity extraction, and summary writing, all designed to enhance your academic pursuits.

 

 

 

Visual Pollution Detection

This system is based on computer vision and deep learning for automatic real-time detection and classification of visual pollution. The system is designed to detect and classify the most common and significant visual pollutants, namely: Bad Billboards, Construction Road Barriers, Damaged Sidewalks, Garbage, Graffiti, and Potholes.

 

 

 

Autonomous UAV Visual Navigation

In this research, the team developed a simulated environment representing the entrance of SDAIA-KFUPM JRCAI and trained a virtual drone to autonomously navigate indoors without collisions using an enhanced deep reinforcement learning model.

 

 

Deep Reinforcement Learning for Object Sorting

Implementation of an adaptive learning scheme based on deep reinforcement learning for performing object sorting in very dynamic and complex environmental scenarios (cluttered object block).

 

Explainable Deep Learning for Pneumonia Prediction

In the demonstration, InceptionV3 and GradCAM were selected before classification and explainability analysis. The results showed that the system accurately classified the uploaded pneumonia X-ray as pneumonia, achieving a 100% accuracy score. Additionally, the GradCAM visualization highlighted the middle, left, and right lung regions, indicating the presence of pneumonia in these areas. When InceptionV3 and GradCAM++ were selected, and a normal chest X-ray was uploaded, the model classified it as normal (no disease) with 77% accuracy. The GradCAM++ heatmap further confirmed the absence of pneumonia, as it focused on the stomach area rather than the lungs.

 

Automated Arabic OCR System

This system focuses on improving Arabic OCR accuracy by using spell-checking correction as a post-processing step. This technique refines OCR output, reducing errors and enhancing overall text recognition quality

 

Salinity Prediction Dashboard

In the demonstration, InceptionV3 and GradCAM were selected before classification and explainability analysis. The results showed that the system accurately classified the uploaded pneumonia X-ray as pneumonia, achieving a 100% accuracy score. Additionally, the GradCAM visualization highlighted the middle, left, and right lung regions, indicating the presence of pneumonia in these areas. When InceptionV3 and GradCAM++ were selected, and a normal chest X-ray was uploaded, the model classified it as normal (no disease) with 77% accuracy. The GradCAM++ heatmap further confirmed the absence of pneumonia, as it focused on the stomach area rather than the lungs.

 

AI-Powered Road Anomaly Detection

This prototype aims to enhance road safety by developing a comprehensive system for detecting, classifying, and reporting road anomalies, such as potholes and cracks (including longitudinal and transverse cracks). The system operates in three modes: Mode-1, where processing is performed on the user’s side; Mode-2, where processing is handled by the roadside unit; and Mode-3, which operates offline.