LOAIS Keeps GBV Guidance Offline and Out of Case Files

LOAIS, an offline AI assistant for gender-based violence work, is being piloted by humanitarian staff in Jordan, Lebanon, Gaza and Syria. The tool answers from approved guidance in English and Arabic, and it is built so a case file never enters the model.

Elly Mulaha, 35, a University of Hertfordshire student who grew up in Kenya, designed it after 14 years in the sector with the United Nations, the Red Cross and USAID-funded programmes. He presented the system on 8 July 2026 at the International Telecommunication Union’s AI for Good session.

LOAIS Answers Questions and Refuses Case Files

The ITU lists the product as the GBViE Localized Offline AI System, a closed offline-first dual-language guidance assistant. Frontline staff, it says, still lack fast, context-specific help on gender-based violence in emergencies, especially where GBVIMS, standard operating procedures and training packs are hard to reach in real time.

LOAIS is a progressive web app. It pairs question-and-answer help with focus-group support tools and safety reminders, all drawn from officially approved inter-agency documents. Those documents include the Inter-Agency Standing Committee guidelines for GBV interventions in emergencies, which set the core actions for cutting risk, reaching care and rebuilding local capacity.

The assistant does not run case management. It does not take incident reports. The ITU line is blunt: all incident-level data remains inside GBVIMS+/Primero, the sector’s existing database for survivor files.

I am passionate about responsible AI and developing practical technologies that help frontline workers make informed decisions in complex situations.

Elly Mulaha, information and knowledge management specialist

That split is the product. A worker can ask how to handle a disclosure, how to run a focus group, or which safety step comes next. The names, dates and case notes stay in the system built for them.

Why the App Installs Under a Neutral Name

A connected chatbot that stores survivor detail is a targeting risk in a crisis zone. The Inter-Agency Standing Committee’s April 2023 operational guidance on data responsibility treats personal data in humanitarian work as a do-no-harm problem, not a software feature. LOAIS is built around that limit.

Personal data is stripped before any AI processing. Collection is anonymous. The install name is a generic “GBViE Tools” profile, so a phone that is searched or seized does not advertise a GBV caseload. A PIN lock and a session that expires on its own close the rest of the gap.

THE SIX BUILT-IN SAFEGUARDS

• PII stripping: Names and other personal identifiers are removed before the model sees a prompt.
• Incident-data blocking: The assistant will not take or store a report that belongs in a case file.
• Off-topic filtering: Questions outside approved GBViE guidance are shut down.
• PIN-lock access: The app opens only with a PIN.
• Session auto-expiry: A forgotten screen does not stay live.
• Neutral install name: The phone shows “GBViE Tools”, not a GBV or AI label.

Those controls make the assistant less capable than a general chatbot on purpose. It cannot browse the open web. It cannot keep a survivor record. It can only replay, in the worker’s language, what the sector has already signed off.

Voice, Arabic Dialects and a Phone With No Signal

The app takes voice and text. It normalises Arabic dialects rather than forcing workers into formal written Arabic, which is a different language from the one used in a camp or a clinic. English sits beside it for staff who switch between the two.

Offline-first is the other constraint. A progressive web app can sit on a phone after it is loaded, then keep answering when the mast is down. That is the point in Gaza and parts of Syria, and it still matters in Jordan and Lebanon when power, fuel or a checkpoint cuts the link.

Mulaha has said he wants practical tools that help frontline staff decide in complex settings, and that the AI for Good invitation was a chance to show how AI can support humanitarian action in a safe and ethical way. The design follows that brief. The model never needs a cloud round-trip to be useful, and it never needs a survivor’s file to answer.

Dr Mohammed Bahja, a senior lecturer in computer science at the University of Hertfordshire, helped lay the technical base. The university’s role is the academic half of a tool whose other half is 14 years of programme work.

Incident Data Stays Inside GBVIMS+

GBVIMS is the sector’s long-running way to classify and count GBV incidents. UNICEF dates the original system to 2006. In 2013 the GBVIMS steering committee joined work on Primero, an open-source GBV case management system that now carries the GBVIMS+ module. Partners on Primero include UNICEF, the IRC, Save the Children, UNFPA, DPKO and the Office of the Special Representative of the Secretary-General for Children and Armed Conflict. The steering group around GBVIMS is the IRC, UNHCR, UNFPA, UNICEF and IMC.

GBVIMS+ does the job LOAIS refuses. It holds digital intake and consent forms, tracks cases, flags work for supervisors and exports aggregate incident statistics. A mobile app is meant to work in low-connectivity settings. It already runs in English, French, Arabic, Spanish and other languages. Access is role-based, so a caseworker sees only their own files.

WHERE EACH SYSTEM STOPS

The GBVIMS site roots that database in the survivor-centred rules: physical safety, confidentiality, respect for the survivor’s wishes, and non-discrimination. It also points to the WHO’s 2008 ethical recommendations on documenting sexual violence in emergencies, and to the 2017 Inter-Agency GBV Case Management Guidelines. LOAIS does not try to replace that stack. It sits beside it as a coaching layer when the manuals are in a pack at the office and the worker is not.

HOW THE DATA STACK WAS BUILT

1. 2006: The original GBVIMS tools are developed for humanitarian settings.
2. 2013: Work begins on Primero, with GBVIMS+ as the GBV module.
3. 2017: The Inter-Agency GBV Case Management Guidelines set the practice GBVIMS+ is built to follow.
4. 8 July 2026: Mulaha presents LOAIS at AI for Good as a guidance assistant that leaves incident data in GBVIMS+/Primero.

Interoperability is the unglamorous test. A second app that secretly starts a case file would duplicate records, split consent and put survivors through the same story twice. LOAIS is designed so that cannot happen inside the model.

The Four Countries Where Guidance Must Work Offline

The pilot list is Jordan, Lebanon, Gaza and Syria. All four already run GBV response through a mix of safe spaces, mobile teams and case management, and all four have stretches where a live connection is a luxury or a risk. The ITU description of LOAIS is written for exactly that gap: low connectivity and high risk, with GBVIMS, SOPs and training materials not easy to pull up on the spot.

In Lebanon, the GBVIMS task force said in its first-quarter 2026 analysis that conflict, movement limits, economic pressure and fear of retaliation still keep survivors from reporting and from services, even when recorded incidents fall. A drop in reports is not a drop in violence. It is often a drop in safe access. A worker in that setting needs a private way to check a procedure without opening a case in the wrong place.

The assistant is for staff, not for survivors tapping a public chatbot. That matters in places with mandatory reporting, weak data laws or active surveillance. A survivor-facing bot that logs a location, a name or a medical detail can hand that trail to someone who should never see it. LOAIS keeps the human in the room and the file in Primero.

None of the public material names the agencies running the four-country pilot, or how many devices are in the field. What is on the record is the operating idea: a closed app, two languages, no case file, and a phone that still works when the network does not.

Fourteen Years of Aid Work Shape the Code

Mulaha is 35 and is studying for an MSc in computer science with artificial intelligence at the University of Hertfordshire. The ITU lists him as an information and knowledge management specialist. The 14-year stretch in UN, Red Cross and USAID-funded programmes is the other credential, and it shows in what he chose not to automate.

Case management is skilled, supervised work. The 2017 guidelines treat it that way. An MSc project that tried to swallow that work would have made a more dramatic demo and a worse field tool. The modest version, a guidance and capacity-building assistant, is the one humanitarian teams will load onto a phone.

Dr Charmagne Barnes, pro vice-chancellor for education and student experience at the University of Hertfordshire, called the project an example of students joining academic learning to professional experience. She also pointed to a commitment to ethical and responsible AI, and to recognition on an international stage. The stage was a 30-minute slot, 14:00 to 14:30 CEST, on 8 July 2026.

The talk does not, by itself, prove the pilot works. It puts a tightly scoped tool in front of the UN tech crowd and leaves the hard test where it belongs, with GBV staff in the four countries.

GBV Tech Projects Keep Hitting the Same Limits

Humanitarian funders have already paid for a wave of GBV tech. Elrha’s GBV Tech Innovation Challenge ran from 2023 to 2025 and backed four projects with grants of up to £150,000: a virtual-reality training tool, a chatbot for community outreach workers who take disclosures, and two projects on remote case management. The lessons from four GBV tech pilots, published on 2 March 2026, start with a warning. New tools are not the win. Safety, language, the real end user and the way systems join up are the win.

WHAT THE ELRHA COHORT HAD TO LEARN

• Start with survivors: Design begins with how people actually reach help, not with a model.
• Name the user early: A staff tool and a survivor tool are different products with different risks.
• Build safety in: Branding, channels and data flows can expose someone before the content does.
• Join existing systems: A parallel case file creates duplicate records and extra burden.
• Test language on site: Localisation, including dialect, is a safety step, not a translation pass.

LOAIS maps onto those lessons with unusual discipline. It is a staff tool. It installs under a bland name. It strips personal data. It speaks Arabic dialects. It leaves GBVIMS+/Primero as the only place a case can live. That is also its ceiling. It will not identify a case, open a referral, or carry a file across a checkpoint. Teams that need those functions still need Primero, paper forms, or a separately designed channel such as Elrha’s South Sudan outreach bot, which sent client data to a remote server and wiped it from the worker’s phone.

A 2025 analysis published by the International Committee of the Red Cross went further on survivor-facing bots. Used for mental health and psychosocial support in conflict, they mostly do not yet line up with GBV guiding principles. One stated risk is a survivor revealing sensitive data to a system that local authorities can reach, including in places with mandatory reporting. LOAIS tries to stay out of that trap by never sitting in front of the survivor and never holding the incident.

The remaining question is operational, not rhetorical. Can a closed, offline coaching layer actually speed a decision in a clinic, a safe space or a mobile team, and will staff trust it enough to keep the PIN on their own phones? The four-country pilot is where that gets answered. Until those results are public, the honest claim is the design claim. The model is small on purpose, the file is elsewhere, and the phone is meant to keep working when the network does not.