Sanjomwa/Civil-Liberties-Knowledge-Assistant

RAG assistant answering questions about internet censorship and digital rights in East Africa (2022–2026), grounded in a curated corpus from OONI, Access Now, CIPESA, and Freedom House — every answer cites its sources. Built for DataTalksClub's LLM Zoomcamp 2026. # Civil Liberties Knowledge Assistant I built a citation-grounded RAG assistant to help researchers, journalists, and civic-tech practitioners investigate internet censorship and digital rights in East Africa (Kenya, Uganda, Tanzania, Ethiopia, Rwanda; 2022–2026). It retrieves and answers from a curated corpus of OONI, Access Now, CIPESA, and Freedom House reports — every answer cites the specific excerpt it draws from, and I made sure thin or single-sourced evidence gets flagged rather than smoothed into a confident-sounding narrative. Built for DataTalksClub's LLM Zoomcamp 2026 capstone project. **Live demo:** app monitoring dashboard ***both on Google Cloud Run, `min-instances=0` — a quiet period means the next request has a few seconds' cold start, not that the link is dead.*** ## Contents - Problem - Demo - Evaluation - Testing - Monitoring - Quickstart - Data and configuration - Deployment - Architecture - Project structure - Decisions and trade-offs - CI/CD - Limitations ## Problem Reporting on internet censorship and digital rights in East Africa is produced by several independent organizations — OONI publishes network measurement data, Access Now tracks shutdowns, CIPESA covers regional policy, Freedom House scores countries annually — each with its own format, scope, and update cadence. A researcher or journalist trying to answer a specific question (did this event actually involve app-level blocking or a full shutdown? which organizations have corroborated a claim, and which haven't?) currently has to manually cross-reference several separate sites and PDF reports, with no single place to ask a direct question and get a sourced answer. That manual cross-referencing has a second, easy-to-miss failure mode beyond being slow: it's easy to accidentally flatten disagreement or thin evidence into a single confident-sounding narrative, because nothing forces the reader to notice when only one organization covers an event, or when two sources …