Project 01AI Digital Library
An Arabic-first storytelling experience designed for safe, engaging, and culturally aligned reading.
Naseej Innovation Lab is the research and development engine of Naseej's product ecosystem. The Lab explores emerging technologies, advances Arabic AI, and transforms research into prototypes, intelligent services, and scalable solutions for education, knowledge, culture, government, and enterprise.
By combining applied research with product development, the Lab helps Naseej evaluate new technologies rigorously, reduce experimentation risk, and accelerate the delivery of AI capabilities that are practical, responsible, and relevant to the region.
To lead Arabic AI innovation by advancing natural language processing, computer vision, and intelligent technologies that transform education and knowledge services across the region.
As the research and development engine of Naseej's product ecosystem, Naseej Innovation Lab develops AI-powered solutions that enhance education and knowledge services, advance Arabic-language technologies, and transform innovative ideas into practical, scalable applications.
The right answers begin with the right questions. We define the problem before selecting the technology.
Creative ideas should strengthen institutions, communities, and the wider knowledge ecosystem.
Change is a fundamental driver of growth, learning, and meaningful transformation.
Great ideas require experimentation, evidence, and the courage to test new possibilities responsibly.
Arabic natural language processing and generative AI
Ethical, trustworthy, and culturally aware AI
AI agents, multi-agent orchestration, and intelligent automation
Semantic search, Retrieval-Augmented Generation, and recommendation systems
Computer vision and human-centered intelligent interaction
Model evaluation, Arabic benchmarks, datasets, and applied AI research
Projects, research, and AI agents from the Lab, brought together in one interactive view. Pick a category, or browse them all.
Project 01An Arabic-first storytelling experience designed for safe, engaging, and culturally aligned reading.
Project 02AI-assisted skills intelligence for more targeted employee development.
Project 03A responsible Arabic AI experience designed to balance helpfulness, safety, and cultural awareness.
Project 05 · NewAn AI-powered experience that presents each artifact’s story according to the visitor’s category through text, voice, and interactive conversation.
A systematic analysis of core Retrieval-Augmented Generation components, with practical guidance for Arabic pipeline configurations.
A systematic evaluation of small language models across Arabic tasks, examining how alignment, instruction adherence, model behavior, and size shape performance.
A suite of agents that helps organize meetings, summarize discussions, and extract decisions, action items, and next steps.
Contact UsAnswers questions about Naseej meetings and retrieves key discussions, decisions, and action items from meeting records.
Contact UsAnswers questions about Naseej products and solutions, including their features, comparisons, and suitability for different needs.
Contact UsReviews presentation content and provides practical feedback to improve its structure, message clarity, and delivery.
Contact UsSummarizes monitored social media events, trends, and mentions into a clear and concise digest.
Contact UsIdentifies the shortest connection path to a target person or company through the user’s professional network.
Contact UsOrganizes initial project information into a structured package covering objectives, scope, stakeholders, risks, and next steps.
Contact Us
An Arabic-first storytelling experience designed for safe, engaging, and culturally aligned reading.
Arabic-speaking families have limited access to digital children's content that combines personalization, accessibility, and clear cultural governance. Many platforms rely on static libraries, while Arabic-first, on-demand storytelling with structured safety controls remains limited.
AI Digital Library generates original Arabic stories tailored to age, interests, reading level, and selected moral themes, with illustrations, narration, diacritics, and comprehension activities in a parent-managed environment.

AI-assisted skills intelligence for more targeted employee development.
Organizations often rely on manual, fragmented, or inconsistent methods to evaluate skills and identify development needs. This can make it difficult to compare current capabilities with job requirements, prioritize training, and track improvement over time.
The system combines employee assessments, manager input, and role requirements to identify skill gaps and recommend targeted learning pathways. AI supports question generation, response analysis, competency comparison, and the preparation of personalized development recommendations.

An AI-powered museum experience that presents each artifact’s story according to the visitor’s category through text, voice, and interactive conversation.
Traditional museum experiences often present the same content to all visitors, despite differences in age, interests, and knowledge levels. Some visitors may also find it difficult to benefit from group explanations or standardized audio narration.
ERTH allows visitors to scan a QR code and select their visitor category. It then adapts the artifact’s content to suit them through personalized text, voice narration, and interactive conversation that enables visitors to ask questions about the artifact.

A responsible Arabic AI experience designed to balance helpfulness, safety, and cultural awareness.
Arabic language models may produce unsafe, culturally inappropriate, incomplete, or excessively restrictive responses when handling complex and sensitive topics. Organizations need models that can remain informative and balanced while applying appropriate safety safeguards.
Noon V2 is an Arabic AI platform built on Gemma 3 12B and specialized through supervised fine-tuning on a culturally aligned Arabic dataset. It is designed to respond to complex Arabic questions in a balanced, culturally aware, and context-sensitive manner while reducing unnecessary refusal and maintaining clear safety boundaries.

A transparent Arabic benchmark for evaluating Small Language Models.
Most widely used language-model benchmarks emphasize English and large-scale models. Arabic-focused evaluation of smaller models remains limited, even though Small Language Models can offer lower cost, reduced latency, improved deployment flexibility, and greater control in resource-constrained environments.
Mizan is an Arabic benchmarking platform that compares Small Language Models across a structured set of Arabic comprehension and generation tasks. An interactive leaderboard helps users examine model quality alongside model size and explore domain-level, skill-level, and failure-pattern results.
To identify effective combinations of Retrieval-Augmented Generation components for Arabic content and provide practical guidance for selecting pipeline configurations across different types of datasets and use cases.
Read the Paper on arXivAcross the evaluated datasets and configurations, sentence-aware chunking achieved the strongest average performance.
BGE-M3 and multilingual-e5-large were among the strongest embedding models in the evaluated settings.
Reranking improved answer faithfulness and relevance, particularly for more complex datasets.
Aya-8B outperformed StableLM-1.6B under the evaluated Arabic RAG configurations.
Contribution: The study provides a systematic comparison of core Arabic RAG components and converts the results into actionable guidance for designing retrieval and generation pipelines.
Practical relevance: Supports evidence-based selection of chunking, embedding, reranking, and generation components.
To evaluate Small Language Models systematically across Arabic tasks and examine how language alignment, instruction-following, model behavior, and size affect performance.
Read the Paper on arXivGemma 3 12B achieved the highest overall score in the evaluated model set.
Aya and C4AI Command Arabic were among the strongest models in the evaluation.
Model size alone did not determine Arabic performance.
Arabic alignment and reliable instruction-following were important performance factors.
Lower-performing models showed recurring issues such as hallucination, language drift, prompt leakage, and incomplete responses.
Contribution: The study introduces a structured Arabic benchmark for Small Language Models and provides model-level, skill-level, domain-level, efficiency, and failure-pattern analysis to support more informed model selection.
Practical relevance: Supports selection of efficient Arabic-capable models for different tasks and deployment conditions.
Specialized digital coworkers that connect AI with knowledge, systems, and workflows. Naseej Innovation Lab develops specialized AI agents and multi-agent workflows that help users retrieve information, prepare decisions, automate repetitive tasks, and interact with enterprise systems through natural language. The goal is not only to answer questions, but also to help users complete work safely, transparently, and efficiently.
The Lab is advancing from isolated, on-demand assistants toward proactive orchestration. In this model, the platform can detect relevant business or service signals, select the appropriate specialized agent, prepare a recommended action, and request user approval before consequential execution.
Identify authorized events, deadlines, changes, risks, or opportunities.
Determine why the signal matters and which agent or workflow is appropriate.
Assemble the required context and create a clear proposed action.
Keep the user or responsible employee in control of consequential actions.
Perform the approved task through secure, role-based integrations.
Record the request, evidence, approval, action, and result for accountability.
Building systems that are useful, trustworthy, culturally relevant, and governed throughout their lifecycle. Responsible AI is integrated into research, model evaluation, product design, and agent deployment: the Lab applies governance according to the risk, audience, data sensitivity, and potential impact of each solution.
Test for misuse, harmful outputs, insecure actions, prompt injection, data leakage, and integration vulnerabilities.
Collect only necessary data, control access, define retention, and protect sensitive institutional and personal information.
Keep qualified people responsible for high-impact decisions, approvals, exceptions, and escalation.
Document intended use, limitations, evaluation results, data sources, model versions, and significant system actions.
The Lab evaluates whether Arabic AI systems understand linguistic variation, maintain Arabic language consistency, respond appropriately to regional contexts, and respect approved cultural and institutional requirements. Cultural alignment is treated as a documented and testable design objective rather than an unlimited claim of correctness.
Strategic priorities that extend the Lab from experimentation to sustained product and regional impact.
Continue developing and evaluating Arabic-first language technologies, datasets, benchmarks, retrieval systems, and models that address the linguistic and operational needs of institutions in the region.
Expand safety evaluation, cultural-alignment testing, explainability, governance, bias analysis, and model-risk management for Arabic AI systems and agents.
Build coordinated agents that detect authorized needs, share structured context, prepare high-value actions, and safely execute approved work across education, knowledge, government, and enterprise systems.
Embed intelligent assistance, semantic search, recommendations, automation, and evidence-linked insights across the Naseej and Medad product ecosystem as governed platform capabilities rather than isolated features.
Create reusable AI services, evaluation frameworks, reference architectures, and product-integration patterns that help successful prototypes move into secure and scalable production environments.
Collaborate with universities, government organizations, technology providers, customers, and research communities on shared challenges, pilots, datasets, evaluation, and knowledge exchange.
Support internships, trainee programs, technical workshops, applied research opportunities, and multidisciplinary projects that build practical capabilities in Arabic AI and responsible innovation.
Publish research papers, benchmark releases, evaluation methodologies, technical insights, datasets, and demonstrations that strengthen Arabic AI research and encourage reproducible progress.
Continue improving the Lab's existing AI projects by expanding validated capabilities, enhancing user experience, strengthening performance and responsible AI safeguards, and preparing selected solutions for broader organizational adoption and scalable deployment. Future development may also include deeper integration with Naseej products and enterprise systems, based on user needs and evaluation results.
From research questions to pilots, products, and measurable operational value.
Joint studies, datasets, benchmarks, publications, and evaluation initiatives.
A focused proof of concept or pilot designed around a measurable institutional challenge.
Arabic capability, RAG, safety, cultural alignment, efficiency, and agent-security evaluation.
Integration of validated AI capabilities into Naseej products, platforms, and customer workflows.
Contact Naseej Innovation Lab to discuss a research collaboration, pilot project, model evaluation, or responsible AI implementation for your organization.
Or contact us at [email protected]