Offline AI tool from Herts student reaches GBV teams in Gaza and Syria

Elly Mulaha’s LOAIS assistant is already guiding humanitarian workers on gender-based violence response in Jordan, Lebanon, Gaza and Syria. The University of Hertfordshire MSc student built the system to run fully offline in Arabic and English, drawing only on approved inter-agency documents.

The 35-year-old Kenyan, who brings 14 years in the sector, presented the tool at the AI for Good Global Summit in July. Pilots are under way with field teams who lack reliable connectivity yet still need fast, safe guidance.

What LOAIS actually does in the field

LOAIS stands for Localized Offline AI System, or more fully the GBViE Localized Offline AI System. It is a closed progressive web application that answers questions, supports focus-group tools and embeds safety reminders. Everything stays grounded in officially approved humanitarian resources.

Workers can type or speak. The system normalises Arabic dialects. It never takes over case management or incident reporting. Those stay inside established platforms such as GBVIMS+ and Primero. Automatic PII stripping happens before any processing. Off-topic queries get filtered. A PIN lock and session auto-expiry add further barriers.

The design choices are deliberate. Frontline staff often cannot reach GBVIMS, standard operating procedures or training packs in real time when the network drops or the risk is high. LOAIS fills that gap without creating a new data honeypot.

Mulaha’s path from UN and Red Cross work to an MSc project

Mulaha grew up in Kenya and spent nearly a decade and a half in information and knowledge management roles. He has worked with the United Nations, Red Cross and Red Crescent Societies, and USAID-funded programmes. His tools of trade already included KoboCollect, ActivityInfo, CommCare, GBVIMS+/Primero and SurveyCTO, plus Microsoft Power Platform and data visualisation.

He is now finishing an MSc in computer science with artificial intelligence at the University of Hertfordshire. He credits senior lecturer Dr Mohammed Bahja for the technical foundations. Pro vice-chancellor Dr Charmagne Barnes called the project “a wonderful example of how students can combine academic learning with professional experience” and noted its ethical focus.

Mulaha himself said he is “passionate about responsible AI and developing practical technologies that help frontline workers make informed decisions in complex situations.” The summit invitation, he added, let him show “how AI can be used to support humanitarian action in a safe, ethical and meaningful way.”

• United Nations roles supporting frontline operations
• Red Cross and Red Crescent Societies
• USAID-funded programmes
• Specialist tools: GBVIMS+/Primero, KoboCollect, CommCare and others

That résumé is why the tool feels built for the job rather than adapted from a consumer chatbot. It is the same reason another European student who built a practical AI tool drew attention: lived constraints produce tighter products.

The larger UNRWA and IFRC project behind the student demo

LOAIS did not appear in isolation. It forms one pillar of a Humanitarian Innovation Programme project led by UNRWA and the International Federation of Red Cross and Red Crescent Societies. The effort targets GBV prevention and response for Palestinian refugees, where rates of violence often rise and only a tiny slice of humanitarian funding reaches the problem.

About 30 percent of women experience physical or sexual violence in their lifetime. In displacement the risks climb further. Frontline teams in the MENA region face poor internet, thin infrastructure and few tools that speak local Arabic dialects. The project therefore pairs community work with men and boys with a technical fix: a voice-based Generative AI technology designed for offline functionality.

Private-sector partners handle model customisation, linguistic tuning and training. Datasets and upskilling cover Jordan, Lebanon, Syria, the West Bank and Gaza. A 2024 UNRWA consultancy listing described LOAIS development as a core deliverable for enhancing first-responder support to survivors.

Why the refusal to connect is the feature

Most commercial AI assumes a cloud link, continuous updates and broad data access. Those assumptions break in a shelter with intermittent power or in an area where traffic analysis itself creates risk. LOAIS inverts the stack. It is an offline-first dual-language progressive web app that keeps every answer inside a closed, pre-approved knowledge base.

That choice has second-order effects. Because the system never accepts incident data, it cannot become a back-door case-management tool or a leak vector. Because it strips PII automatically, even anonymous usage logs stay clean. Because it filters off-topic questions, it resists the drift that turns open models into liability. Field teams already tinker with local models for exactly these privacy and latency reasons; LOAIS simply ships the hardened version under institutional cover.

Mulaha’s speaker profile at AI for Good notes his nearly 14 years of experience driving data-led innovation across complex emergencies. The tool is the product of that time spent watching what fails when the network dies.

What the system will not do

LOAIS deliberately stops short of several tempting capabilities. It does not store or process individual survivor stories. It does not replace referral pathways or clinical judgement. It does not claim to predict risk scores or automate decisions. Those boundaries keep it inside the “guidance and capacity-building” lane rather than the high-stakes case-management lane already covered by GBVIMS+.

What we know

• Pilots are live with humanitarian workers in four Middle East locations
• Content is limited to approved inter-agency GBViE documents
• Safeguards include PII stripping, PIN access and session expiry

What remains open

• Scale of current pilot user numbers and measured time savings
• Long-term ownership and update process after the student project ends
• Exact private-sector partners handling model customisation

Those limits are features in a sector that has learned hard lessons about data sovereignty and secondary trauma. An always-connected model that “helps” by scraping case notes would create more problems than it solved.

How the university framed the achievement

University of Hertfordshire statements emphasise the combination of classroom technique and prior professional years. Barnes highlighted both the technical innovation and the “strong commitment to ethical and responsible AI.” The institution presents Mulaha as one of many students turning study time into field-ready work.

The timing of the AI for Good slot on 8 July 2026, a 30-minute session, placed the demo in front of the global audience that cares about precisely these constraints. The event page describes the persistent operational gap LOAIS targets: staff who cannot pull SOPs or training materials when they need them most.

Field reality still sets the pace

Connectivity blackouts, dialect variation and the absolute need to keep survivor data offline are not edge cases in the places LOAIS now runs. They are the daily environment. By accepting those conditions as non-negotiable, Mulaha produced a tool that can actually be installed and trusted. The pilots will show whether the guidance is fast enough and accurate enough under stress. For now the architecture itself is the proof of concept: responsible humanitarian AI looks less like a general chatbot and more like a locked briefcase of approved answers that works when the lights go out.