ReliefWeb is a specialized digital service provided by OCHA, and one of the humanitarian community's most widely used information platforms. Since 1996 it has aggregated, curated and disseminated critical information from more than 2,000 organizations across more than 200 countries and territories. It publishes over 50,000 reports annually, receives about 15 million unique visitors a year and its API handles more than 400 million calls per year from organizations integrating humanitarian data into their own platforms.
ReliefWeb recently wrapped up a productive pilot with Dataminr's AI for Good program. While the Humanitarian Reset and internal organisational change meant we couldn't complete everything we set out to do, we made real progress worth sharing.
Why We Started
The humanitarian sector is undergoing much-needed change. More responses are being led by local actors, which means a rise in information produced in local languages that have historically been under-represented on global platforms like ReliefWeb. For a platform built around making critical information findable and usable, that gap matters. A significant share of high-priority reporting in languages such as Arabic, Portuguese and Ukrainian remains unshared with platforms like ReliefWeb.
AI-assisted translation offered a way to change that equation without placing new burdens on partners or overwhelming the ReliefWeb editorial team.
We also knew from conversations with partners that the figures buried inside humanitarian reports (people affected, funding levels, displacement numbers) were valuable data points that required manual effort to surface. Both problems felt like good candidates for AI-assisted solutions.
What Dataminr's AI for Good Program Made Possible
Dataminr's AI for Good program connects NGOs and UN agencies with Dataminr AI scientists to apply AI in humanitarian and human rights contexts. The program gave us access to technical expertise we couldn't have replicated internally, and a structured way to move from idea to prototype quickly.
What We Built
We reached the proof-of-concept stage on two workstreams.
The first was automated translation. We demonstrated that ReliefWeb documents arriving in Arabic, Portuguese and Ukrainian could be translated into English automatically, making them discoverable and usable for a wider audience without waiting for manual translation.
The second was structured data extraction. We did foundational work on pulling key figures from unstructured report text (numbers like people affected or funding totals) and surfacing them as dedicated metadata with API endpoints. The goal was to make those figures consumable by other services without anyone having to read the full document first.
Alongside the working code, the Dataminr team produced a guidance document on LLM-based translation covering the approach, considerations and lessons learned, as well as a GitHub repository with documentation and code for both the translation and key figure extraction work. Neither feature reached production at this time. Staffing changes and a broader organizational transition meant we had to stop before completing testing and rollout.
What We Learned
Short, scoped pilots like this are valuable even when they don't cross the finish line. We now have working code, a clearer picture of the technical requirements and a better sense of where the real complexity lies. The key figures extraction work surfaced harder questions about consistency and validation that would need to be resolved before any production deployment.
What Comes Next
We're sharing this now because the problems we were trying to solve haven't gone away. Language barriers and unstructured data remain real friction points for humanitarian information users. We hope this write-up is useful to others thinking about similar challenges, and we remain open to continuing this work if the right opportunity arises.
We're grateful to the Dataminr AI for Good team for their partnership and for making this kind of exploratory collaboration possible.