01 · Public-discourse intelligence
Vigil
See the conversation before it shapes the election.
Reads Ghana's public discourse across news, social media and live radio, runs an LLM intelligence layer over it, and sets the result against electoral and socioeconomic data to forecast outcomes.
- 01Many-source ingest — 24 news-site scrapers and 9 platform adapters (X, Facebook, YouTube, TikTok, Telegram, Bluesky and more).
- 02Radio to text — ffmpeg stream capture, Whisper transcription and an ad filter before analysis.
- 03LLM intelligence layer — sentiment, entities, topics and emotion, with backend failover from vLLM to Anthropic to Ollama.
- 04Trained forecasts — a constituency-level vote-share ensemble (Ridge and gradient boosting) with a 2024 holdout back-test built in.
- 05Live-editable schedule — a database-driven Celery scheduler: edit a cron row and the workers follow within 30 s, no restart.
- 06Coordination and anomaly — near-duplicate posts across authors in a time window; deviations from each profile's baseline.
- 07Observability and analytics — Grafana dashboards for the pipeline, errors and cost, and Superset over the reporting schema (on a branch, not yet on the main line).
Powered byCweCi
- 24 news scrapers
- 9 platform adapters
- 14 services
- 680+ tests
SurfacesWebSelf-hostedAPI
- Python · Celery
- Redis
- PostgreSQL
- Solr
- vLLM · Claude · Ollama
- Whisper
- scikit-learn
- FastAPI
- Next.js