A media-monitoring service that tracks mentions of a topic across the web, scores their sentiment and emotions via an LLM, and turns the results into trends and reports.
| Repository | Stack | Role |
|---|---|---|
| backend | FastAPI, dishka (DI), PostgreSQL, LangChain | REST API, data collection, LLM analysis |
| frontend | React 18, TypeScript, Vite, Tailwind CSS, Recharts | Web UI: topics, sources, analysis charts |
| infra | Traefik, PostgreSQL, Docker Compose | Reverse proxy (TLS via Let's Encrypt) and database |
Backend is layered api → services → repositories → db, with a separate
digital-barometer-db package (SQLAlchemy models + Alembic migrations)
shared across services.
Core: FastAPI, dishka (DI), Pydantic Settings
Data sources: GDELT Doc API, NewsAPI, RSS feeds, Google Trends (via SerpApi) —
pluggable through a ConnectorFactory, fetched concurrently with a
configurable outbound proxy
AI: LangChain, OpenAI-compatible LLM endpoint — batched sentiment and emotion scoring, topic summaries; falls back to a regex-based heuristic when the LLM is unavailable
Database: PostgreSQL + SQLAlchemy (async) + Alembic
Infrastructure: Docker Compose, Traefik (automatic TLS), GitLab CI
(test → build → deploy, staging + production)
Sensitive data (API keys, Authorization headers) is redacted from error
logs before they're persisted.
| Table | Purpose |
|---|---|
topics |
Monitored topics and their keywords |
sources |
Configured data sources (GDELT, NewsAPI, RSS, Trends) per topic |
analysis_runs |
A single analysis execution for a topic over a date range |
source_results |
Per-source fetch outcome within a run (status, raw payload, item counts) |
mentions |
Individual mentions collected from sources, with sentiment/emotion scores |
trend_points |
Time-series metrics per source (e.g. Google Trends values) |
analysis_metrics |
Aggregated sentiment/emotion counts and the resulting "barometer" score |
reports |
Generated report files per analysis run |
Sources are fetched concurrently and normalized into a common model
(Mention / TrendPoint / SourceResult), deduplicated by a SHA-256
content hash. If the LLM is unavailable, sentiment/emotion scoring falls
back to a regex-based heuristic instead of failing the run. Each source
result is tracked independently, so a run can finish as success,
partial, or failed depending on which sources succeeded.
See each repository's README for local setup and CI/CD details.
