Our work

Millions of daily social-media and public data points made early threats hard to spot — and policing stayed reactive, relying on post-incident reports.

• AI horizon-scanning monitored and filtered open-source platforms in real time
• Integrated with police databases to predict crime hotspots dynamically
• Early-warning signals enabled proactive patrol deployment

Dangerous near-misses on roads often go unrecorded until a crash happens.

• Used real-time video analytics to spot near-misses
• Sent proactive road-safety alerts
• Gave planners the data to fix risky spots

Diagnosing Mycetoma relied on specialised manual reading of lab slides; limited expert access slowed response and raised misclassification risk.

• Combined computer vision and LLMs to classify, segment and assess pathology slides
• Built intelligent support tools for physicians and research teams
• Enabled chatbot-assisted documentation and data standardisation

“See Also” sections were manually curated and often outdated, so readers lacked links between related articles.

• Built an NLP system using BERT to generate semantic vectors for articles
• Used cosine similarity and article features to recommend related links
• Deployed on selected pages and presented at WikiArabia 2024, Oman

Students struggled with article selection, structuring and neutral writing, and educators lacked guidance on using AI in real tasks.

• Integrated generative AI for brainstorming, outlining and multilingual drafting
• Created AI-assisted workflows aligned with Wikipedia standards
• Presented at EduWiki 2025 in Colombia

Most voice AI struggles with the many dialects of Arabic.

• Built voice AI that understands Arabic dialects
• Deployed across banking, health and government
• Tuned for local language and culture

Search pipelines were slow, compliance reporting was manual, and outsourced chatbots lacked contextual understanding.

• Built an advanced RAG pipeline with hybrid search and reranker integration
• Developed a full-stack FastAPI, PostgreSQL and Gradio platform to triangulate AI insights
• Migrated chatbots in-house with Mixtral fine-tuning and dynamic tagging

Government systems struggled to parse structured legal documents at scale; token-classification models lacked precision and high compute limited adoption.

• Fine-tuned Hugging Face models using LoRA, pruning and ONNX optimisation
• Added layout-aware computer vision for form and signature parsing
• Deployed with ONNX Runtime for real-time inference in production

Creating educational or storytelling video required many manual tools, and text-to-video workflows were inconsistent and resource-heavy.

• Built a generative pipeline using Flux Stable Diffusion with LoRA adapters
• Automated video creation from text — voice, character and scene synthesis
• Published as peer-reviewed research at NAACL 2025

HR teams faced high walk-in volumes and repetitive queries, and staff spent significant time on FAQs instead of strategic work.

• Deployed a retrieval-augmented chatbot to automate common HR responses
• Integrated with internal databases for accurate policy access
• Enabled round-the-clock multilingual support on mobile and desktop