mellowfm.com – Read, Search and Check News

What it does

mellowfm.com gathers reporting from many news outlets, mostly Ghanaian, into one archive. Readers can catch up on the news, search it, and check a claim against what was actually reported. The site opens on the Ghana edition by default.

Who it's for:

• Readers, mostly on phones, who want a quick, calm way to follow the news and check something they heard
• Admins, who submit, approve and monitor news sources and crawls

What readers can do (no account needed):

• Browse today's news by edition (Ghana, worldwide or another country) and by topic
• Search the archive in plain language
• Check a claim against the archive's own stored reporting
• Open articles in a clean reader page, with the publisher credited and linked
• Read topic briefs (Politics, Business & Economy, Sport, Tech, Health, Entertainment, General), also available as RSS
• Subscribe to briefs by email, with double opt-in and one-click unsubscribe

What makes it different:

• Verdicts come from the archive itself. A claim is judged by six separate checks: entailment, contradiction, timing, source authority, numbers and entity consistency. Every verdict links to the articles behind it, and "inconclusive" is a real outcome.
• Briefs rank stories by independent coverage. Articles about the same event are merged into one story. Stories are scored on how many outlets reported them (45%), freshness (25%), source trust (15%) and reporting speed (15%).
• Publishers are credited everywhere their work appears.

WhatsApp sign-up exists, but WhatsApp sending isn't connected yet. Its only provider is a console logger that sends nothing.

How I built it

Frontend

• React 19, Vite 7, Tailwind CSS 4 and React Router 7
• Two separate apps, a public site and an admin, that share one code folder
• Apple-style design language, with a dedicated accessibility pass
• It is a client-rendered app, so the API serves crawler-readable versions of each public page, with title, canonical tag, Open Graph tags and JSON-LD. nginx rewrites public routes to these. The API also serves sitemap.xml, robots.txt and llms.txt.
• Article reader pages are noindex, follow, since they republish other outlets' text

Backend

• Python 3.11 with FastAPI, async SQLAlchemy 2, asyncpg and Pydantic 2, about 27,000 lines across 175 modules
• Search and Check a claim are public and rate-limited. Admin routes use an API key or an approved Google sign-in.

Ingestion

• Crawling uses crawl4ai (headless Chromium), sitemap and RSS checks, and trafilatura for text extraction
• Celery with Redis runs separate workers for crawling, RSS, RSS scheduling and briefs, plus one beat scheduler
• A quality gate rejects error pages, listing pages and boilerplate. Raw pages are kept, so articles can be re-extracted safely.
• Each publication date records where it came from: page metadata, RSS, URL or a backfill

Search and claim checking

• PostgreSQL with pgvector, plus Postgres full-text search. Results are merged by reciprocal-rank fusion, then reranked.
• Embeddings use all-MiniLM-L6-v2 and reranking uses ms-marco-MiniLM-L-6-v2. Both are baked into the image and run offline.
• The Search page uses no LLM
• Claim checking uses DSPy modules and a local Phi-4 Mini model (4-bit) served by Ollama. Production runs it on a CPU-only server. Hosted providers can be switched in by configuration.
• Brief summaries are written in the background, so briefs are saved first and summaries fill in afterwards

Data and infrastructure

• PostgreSQL behind PgBouncer in transaction-pooling mode, and Redis for cache, Celery broker and results
• Alembic migrations, applied by hand only (see "What I learned")
• Prometheus metrics with alert rules, structured logs, optional OpenTelemetry tracing, and Flower for Celery
• Sentry error tracking, switched on only when a SENTRY_DSN is set
• Search and claim events are logged for offline evaluation, without IP addresses, API keys or user identity. The log can be turned off.

Testing

• About 600 backend test functions across 95 test files, covering ingestion, date handling, briefs, country ranking, auth, SEO and migrations
• End-to-end tests against a populated database exist but are excluded from CI
• Frontend testing is thin: one test file for the shared helpers