Building the Foundations for Àdàbà AI: From Legal Frameworks to a Working Prototype
August 2026 Progress Update
What should an AI system rely on when it is being built to support survivors of technology-facilitated gender-based violence?
How should sensitive survivor information be protected? How can an AI system produce useful legal outputs without inventing law? And what should the experience look like for a survivor compared with the lawyer who eventually reviews their case?
These questions shaped much of our work on Àdàbà AI throughout August.
Àdàbà AI is being developed as a smart digital tool to help survivors of technology-facilitated gender-based violence document incidents, organize evidence and move more efficiently towards legal support. The project is designed around a simple principle: AI should support legal and survivor-support professionals, not replace them.
August was an important month because the project began moving beyond high-level design into the foundations that will make a working prototype possible.
Building the legal and data protection foundations
Before an AI system can responsibly handle sensitive survivor information, the rules governing that information need to be clear.
During August, the team worked on a survivor-centric data protection framework, alongside draft terms and conditions for users of Àdàbà AI. The intention is not simply to reproduce generic technology policies, but to develop documentation that reflects feminist and community-centred approaches to privacy, consent and survivor safety.
The team also identified additional documentation needed to support the system, including intake forms and criteria for assessing the strength and relevance of cases.
A recurring principle was accessibility. Terms and conditions should not read like the legal documents of a large software vendor. Survivors need to understand what they are agreeing to, what information is being collected and how that information will be handled.
These documents are therefore being treated as part of the product itself, rather than administrative paperwork added at the end. For any CSO building a survivor-facing tool, this is worth borrowing directly: write consent and data-use terms as UX work, with the same review rigor as a wireframe, not as legal boilerplate bolted on before launch.
Preparing the knowledge base the AI will rely on
One of the biggest pieces of work in August was deciding what information should form the foundation of Àdàbà AI’s legal knowledge base.
The team began organizing relevant Nigerian and ECOWAS laws, court decisions, legal guidance and related materials into a structured corpus. The materials are categorized into chapters and tagged with information such as jurisdiction, document type, legal category, source, date, confidentiality and applicable law.
This structure matters because the AI should not simply be asked to “know the law.” It needs to retrieve the right legal materials for a particular situation.
The team also discussed the value of including cases that did not result in formal judgments. Many gender-based violence incidents do not reach appellate courts, while others are withdrawn or abandoned. These records can still provide useful context about real-world experiences and legal processes, provided they are clearly distinguished from adjudicated cases.
Legal drafting guidance is also being considered as part of the corpus and system instructions. Templates and quality standards can help the system produce more structured first drafts while keeping lawyers responsible for reviewing and validating the final legal output.
Choosing Retrieval-Augmented Generation over fine-tuning
A major technical decision in August was to use a Retrieval-Augmented Generation (RAG) approach rather than begin by fine-tuning a model.
The reason is practical as well as technical. RAG allows Àdàbà AI to work with a curated knowledge base without requiring the project to retrain a large model on thousands of examples — a resource commitment most CSOs and WROs building similar tools on limited budgets simply don’t have.
When a case narrative is submitted, the system can identify relevant facts, dates, actors and events, then use those details to retrieve relevant material from the knowledge base. A language model can then generate an output grounded in those retrieved sources.
The planned architecture combines RAG with rule-based controls designed to constrain how the language model responds. The goal is not to make the AI more creative. It is to make it more controlled.
Potential outputs include incident summaries, timelines, legal categories, evidence summaries and draft legal documents that can then be reviewed by a qualified human.
This architecture also creates room for the knowledge base to evolve. New materials can be added, outdated materials can be identified and different jurisdictions can eventually be supported without rebuilding the entire system from scratch. This is the second reusable lesson from August: RAG isn’t just a technical shortcut — for a resource-constrained organisation, it’s the difference between a system you can realistically maintain and one that requires a retraining budget you don’t have.
Designing security around the survivor, not around the database
Because Àdàbà AI will deal with potentially sensitive evidence, security has been considered alongside the architecture rather than after it.
The planned system includes role-based access controls and case-level permissions so that information is only available to authorized users.
Uploaded evidence is intended to sit in a secure evidence vault rather than ordinary queryable storage. Hashes, timestamps and metadata will be recorded to support evidence integrity and auditing, while encryption will protect information both in transit and at rest.
The team also discussed AI-specific security risks, including controlling what information can be sent to external language-model providers and protecting the system against prompt injection.
Another important decision concerned court documents. Although some court documents are public records, making them broadly accessible online can create additional risks for survivors whose identities or personal information appear in those records. The team therefore agreed that relevant materials should be redacted before inclusion where necessary.
This reflects a broader principle guiding the project: the fact that information is technically accessible does not automatically mean it should be made freely searchable inside an AI system.
Moving from wireframes to a product people can actually use
August also saw significant progress on the user experience.
Low-fidelity wireframes were refined while the design team began developing high-fidelity versions. By the end of the month, the team had begun thinking more concretely about two different experiences: one for survivors and another for legal professionals.
For survivors, the interface is being designed to feel calm and less intimidating. Soft visual elements and the Àdàbà dove symbol are intended to create a sense of safety, while the language emphasizes control, privacy and access to human review.
The reporting flow is designed around flexibility. A survivor will be able to create an account, save their progress and return later rather than being forced to complete a sensitive report in one sitting. Evidence upload can be skipped and added later. The system is also being designed to allow survivors to type or speak their testimony. An emergency support option is planned to be available before personal information is entered.
The lawyer-facing experience is intentionally different. It will have a more professional interface designed around reviewing cases, evidence and legal information while maintaining a consistent Àdàbà identity.
Evidence needs more than a file upload
The August design discussions also pushed the project to think more carefully about what “evidence” means.
Àdàbà AI’s evidence vault is being designed to give uploaded materials different statuses, including Accepted, Under Review and Needs Context. This creates a feedback loop between survivors and lawyers — a lawyer can flag an item that requires clarification rather than simply treating the evidence as complete or unusable.
The team is also exploring how metadata can support evidence verification. Uploaded files can be hashed to help establish whether a file has changed. However, this does not automatically establish when the event shown in that file actually happened. Social-media metadata and timestamps may provide additional context, but accessing this information can involve technical, legal and consent-related challenges.
WhatsApp presents a particular challenge because of its restricted APIs and the sensitivity of the information involved. Discussions with technical partners are underway to better understand what may be feasible.
This is an important distinction for the project: technical evidence of file integrity is not the same thing as proof of an event’s authenticity or timing.
Learning responsible AI through the GAIC bootcamp
August also gave the team an opportunity to deepen its practical understanding of how responsible AI systems are built, through the Gender and AI Innovation Collective (GAIC) bootcamp.
The training covered data audits, data cleaning, chunking, embeddings, RAG systems, prompting, guardrails and model selection. The team also experimented with an AI sandbox environment, using legal materials and other datasets to understand how knowledge bases behave in practice.
One lesson was particularly relevant to Àdàbà AI: the quality of an AI system depends heavily on the quality of the data and instructions surrounding the model. Poorly labelled documents can produce poor retrieval. Badly structured information can result in incomplete answers. Outdated information can undermine otherwise sophisticated systems.
The training therefore reinforced the importance of treating data preparation as core product development work rather than a technical task that happens in the background.
The bootcamp also brought localization into sharper focus. AI systems intended for African contexts cannot assume that language, literacy, culture and user expectations will be the same everywhere. While Àdàbà AI’s initial knowledge base will begin in English, the architecture is being designed with future regional adaptation in mind.
What we have not solved yet
Responsible development also means being honest about what remains unfinished.
Survivor consultation had not yet taken place by the end of August.
The team has faced understandable challenges in engaging survivors, including concerns around privacy, ongoing harassment, trauma and personal safety. While lawyers have already been involved in reviewing aspects of the project, survivor feedback on the prototype remains an important next step.
We are therefore not claiming that survivor feedback has already been integrated into the current prototype.
Instead, the team is preparing for a dedicated consultation on 11 September, with a target of engaging 10 survivors to review the low- and high-fidelity designs. The aim is to gather practical feedback on the reporting experience, language, evidence process and other aspects of the product.
This distinction matters. Building a survivor-centred tool requires more than designing what appears to be a survivor-friendly interface. The people who may ultimately use it need meaningful opportunities to shape how it works.
Why this matters beyond one organisation
Most early-stage legal-AI projects in under-resourced settings face the same temptation: build fast, fine-tune a model on whatever data is available, and treat consent documentation as a formality to finalize before launch. August’s decisions on Àdàbà AI were, in effect, a rejection of that path on three fronts — choosing RAG over fine-tuning to stay maintainable on a modest budget, treating data-protection documentation as product design rather than legal paperwork, and refusing to claim survivor input that hasn’t actually happened yet.
None of these are dramatic decisions individually. Together, they are what separate a prototype that only demonstrates AI can generate text from one that could be responsibly adopted by another under-resourced WRO facing the same constraints DigiCivic Initiative is working in.
From foundations to a working prototype
By the end of August, Àdàbà AI had entered a more active development phase.
The backend team had begun work on the database architecture using PostgreSQL, with authentication identified as the first foundational feature. Survivor story capture and abuse-reporting workflows are next in the development sequence, followed by AI integration.
Frontend development has also started, with the team moving towards incremental deployment so that components can be tested and improved as they are built.
At the same time, the legal corpus is being organized, the RAG architecture is being developed, security controls are being defined and the user interfaces are moving from wireframes towards a working product.
These activities are happening in parallel because the project has a clear milestone ahead: a pilot target of 31 October 2026.
The work supported through this funding is allowing a small multidisciplinary team to bring together legal research, survivor-centred design, data governance and technical development in one process. The immediate objective is not simply to demonstrate that AI can generate text. It is to determine whether AI can be responsibly integrated into a sensitive legal-support workflow.
Looking ahead
September will be about testing what we have built against the people and requirements it is intended to serve. The immediate priorities are to:
- Complete and refine the legal and data protection documentation
- Continue building and structuring the legal knowledge corpus
- Incorporate legal stakeholder feedback into the product design
- Conduct the planned survivor consultation on 11 September
- Continue backend and frontend development
- Test the RAG architecture and evidence workflows
- Explore the feasibility of technical integrations, including WhatsApp metadata retrieval
- Use feedback from survivors, lawyers and technical stakeholders to refine the prototype
By the end of the funding period, success should not simply mean that Àdàbà AI works technically. It should mean that a survivor can use the system without being placed at greater risk; that a lawyer can receive a clearer, better-organized case file; and that the AI’s outputs remain grounded in identifiable legal sources and subject to human review.
That is the standard we are building towards.
Àdàbà AI is still a prototype. But in August, the foundations for testing that standard became much more concrete.
Àdàbà AI is developed by DigiCivic Initiative as part of the A+ Alliance Gender & AI Innovation Collective, supported by Code for Africa. This post is published under a Creative Commons Attribution 4.0 International licence (CC BY 4.0) — reuse and adaptation are encouraged with attribution.