An open-source rumor-triage pipeline for the 2026 Ebola Bundibugyo response in eastern DRC & Uganda. It captures what communities are really saying, clusters scattered reports into thematic threats with an LLM, and drafts accurate, culturally sensitive counter-messaging for frontline teams.
The Bundibugyo strain has no widely approved vaccine or specific therapeutic, so control depends on trust: early detection, safe care-seeking, contact tracing, and dignified burials. Fake cures, resistance to isolation, and rumors about treatment centers can undermine containment within hours.
An open-source pipeline on top of KoboToolbox. It ingests community reports, uses an LLM to cluster scattered rumors into high-level thematic threats, and instantly generates targeted counter-messaging aligned with official public-health guidance — with a human always in the loop.
Responders, community leaders & field workers log rumors via a simple KoboToolbox form — web or offline KoboCollect.
An LLM groups dozens of localized reports — e.g. "poisoned food at the clinic" — into single semantic threat themes.
It auto-generates localized, de-escalating communication briefs optimized for radio, WhatsApp & CHV talking points.
Trained moderators & RCCE focal points review every output. The AI drafts and analyzes — it never decides.
KoboToolbox form with text-entry fields. Works in any browser or offline via KoboCollect.
Custom backend pulls submissions via the Kobo REST API (polling) or Webhooks (real-time push).
Text routed to an LLM via OpenRouter — swap freely between Llama 3, Claude or Gemini.
Thematic clustering of rumors plus auto-drafted, de-escalating counter-messages.
A simple frontend surfaces active rumor trends and ready-to-use briefs for RCCE teams.
WhatsApp dominates communication across the region but is encrypted and private — so the system never monitors it directly. Instead it draws on opt-in tips, public posts, and radio. Radio still reaches an estimated 80%+ of Ugandans and broadcasts in local languages, making it more influential than social media in rural areas.
Community volunteers forward rumors they see in WhatsApp groups — an Uchaguzi/Ushahidi-style crowdsourced model that respects encryption.
Africa's largest fact-checking organization (Johannesburg, Nigeria, Senegal, Kenya). A reference partner for verified information rather than a build-on platform.
The strongest systems fuse radio transcription, human WhatsApp monitors, public-API social scraping, and SMS — no single angle catches everything.
Stand up a KoboToolbox deployment and design a simple form with text-entry fields for rapid rumor reporting.
A lightweight service (Node.js, Python, Go — any stack) retrieving incoming reports via the Kobo API or Webhooks.
Connect the stream to an LLM (via OpenRouter or similar) to parse, categorize and synthesize text into narratives and counter-messages.
A simple open-source frontend (React, plain HTML, or Streamlit) showing active rumor trends and drafted counter-messaging.
Existing transcription tools handle Swahili and Luganda, but local dialects aren't supported — a local translator on the team is needed. We plan to validate in one mostly-Luganda-speaking community (the Baganda) first.
Encryption means no scraping. Rumors spread fastest there in urban areas, so coverage depends on opt-in tips, a Business API number, and human monitors.
This project is grounded in official public-health reporting and humanitarian situation updates. Key live trackers and references below.
Disease Outbreak News — Bundibugyo, DRC & Uganda.
Ebola Disease: current situation summary.
Official Ebola Bundibugyo 2026 dashboard.
Institute of Public Health epidemic dashboard.