Presidential-level intelligence briefings, built with AI, tailored for you.
Meridian cuts through news noise by scraping hundreds of sources, analyzing stories with AI, and delivering concise, personalized daily briefs.
Presidents get tailored daily intelligence briefs. Now with AI, you can too. Meridian delivers:
- Key global events filtered by relevance
- Context and underlying drivers
- Analysis of implications
- Open-source transparency
Built for the curious who want depth beyond headlines without the time sink.
- Source Coverage: Hundreds of diverse news sources
- AI Analysis: Multi-stage LLM processing (Gemini) for article and cluster analysis
- Smart Clustering: Embeddings + UMAP + HDBSCAN to group related articles
- Personalized Briefing: Daily brief with analytical voice and continuity tracking
- Web Interface: Clean Nuxt 3 frontend
graph TD
A[RSS Feed URLs] --> B(Scraper Workflow CF);
B --> C[Article Metadata DB];
C --> D(Article Processor Workflow CF);
D -- Fetches --> E{Content Extraction};
E -- Standard --> F[Direct Fetch];
E -- Complex/Paywall --> G[Browser Rendering API];
F --> H[LLM Article Analysis];
G --> H;
H --> I[Processed Articles DB];
I --> J(Brief Generation Python);
J -- Embeddings --> K[UMAP/HDBSCAN Clustering];
K --> L[LLM Cluster Review];
L --> M[LLM Deep Analysis JSON];
M --> N[Markdown Summary Generation];
O[Previous Day TLDR DB] --> P{Final Briefing LLM};
N --> P;
P --> Q[Final Brief Markdown];
Q --> R[Reports DB];
R --> S(Frontend API CF);
S --> T[Frontend UI Nuxt];
- Scraping: Cloudflare Workers fetch RSS feeds, store metadata
- Processing: Extract text, analyze with Gemini for relevance and structure
- Brief Generation: Cluster articles, generate analysis, synthesize final brief
- Frontend: Display briefs via Nuxt/Cloudflare
- Infrastructure: Turborepo, Cloudflare (Workers, Workflows, Pages)
- Backend: Hono, TypeScript, PostgreSQL, Drizzle
- AI/ML: Gemini models, multilingual-e5-small embeddings, UMAP, HDBSCAN
- Frontend: Nuxt 3, Vue 3, Tailwind
Prerequisites: Node.js v22+, pnpm v9.15+, Python 3.10+, PostgreSQL, Cloudflare account, Google AI API key
git clone https://github.com/iliane5/meridian.git
cd meridian
pnpm install
# Configure .env files
pnpm --filter @meridian/database db:migrate
# Deploy via Wrangler, run Python briefing notebook manually- ✅ Core Pipeline: Scraping, processing, analysis working
- ⏳ Top Priority: Automate brief generation (currently manual Python notebook)
⚠️ Monitoring: Improve scraping robustness- 🔜 Future: Add testing, newsletter distribution
This project benefited significantly from AI assistance:
- Claude 3.7 Sonnet: Contributed to early architecture brainstorming, generated browser js scraping scripts, refined prompts, and called me out when I was overthinking or overengineering.
- Gemini 2.5 Pro: Excelled with long-context tasks - comparing outputs across different prompt variants, reviewing the entire codebase before opensourcing, and nailing the analytical tone for briefs
- Gemini 2.0 Flash: The true unsung hero of this project - blazing fast, dirt cheap, and surprisingly capable when prompted well. It's the workhorse that makes running meridian economically viable without sponsors or grants. Essentially free intelligence at scale.
The first two compressed months of dev work into days and made building this way more fun. But Flash isn't just a time-saver—it's the engine that makes Meridian possible at all. No human is reading 2000+ articles daily and analyzing 100+ story clusters. Having AI peers for brainstorming felt like cheating; having AI workers for the actual intelligence pipeline feels like living in the future.
MIT License - See LICENSE file for details.
Built because we live in an age of magic, and we keep forgetting to use it.