Meridian

AI knowledge graph that extracts concepts from your notes, maps them visually, and finds contradictions and unexpected connections across everything you've ever added. Your ideas, challenged by AI.

Comment

Inspiration

I kept running into the same problem: I read philosophy, study neuroscience, train martial arts, and think about startups, but no tool could tell me when ideas from one domain contradicted or connected to ideas from another. ChatGPT forgets what you said last month. Note apps store things but never challenge them. I wanted something that reads everything I have ever added and argues with it.

What it does

Meridian is an AI-powered personal knowledge graph. You paste in text, notes, or ideas and it extracts key concepts, maps them as an interactive visual graph, and automatically finds contradictions and unexpected connections across your entire knowledge base.

For example: paste in Stoic philosophy about dissolving the ego, then paste in notes about how ego drives competitive performance in BJJ. Meridian flags the contradiction and explains the tension between them.

How we built it

  • Frontend: React 18 + TypeScript + Tailwind CSS with react-force-graph-2d for interactive graph visualization, deployed on Vercel
  • Backend: Python + FastAPI, Dockerized for Google Cloud Run
  • AI: Google Gemini 2.5 Flash for concept extraction, contradiction detection, and connection surfacing
  • ML: sentence-transformers/all-MiniLM-L6-v2 from HuggingFace, running locally in the Python backend, generating 384-dimensional embeddings for semantic similarity
  • Database: MongoDB Atlas with Atlas Vector Search for fast cosine similarity queries across all stored concepts

The ingestion pipeline works like this: text comes in, Gemini extracts structured concepts and relationships, MiniLM generates embeddings, MongoDB stores everything, Vector Search finds similar existing concepts, and Gemini analyzes pairs for contradictions and connections.

Challenges we ran into

Building the full pipeline (text to concepts to embeddings to vector search to contradiction detection) end to end solo in 24 hours. Debugging MongoDB Atlas Vector Search filtering. Managing Gemini API rate limits and billing across multiple Google Cloud projects.

What we learned

How to build and query vector search indexes, how to chain LLM reasoning with embedding models for hybrid AI pipelines, and how powerful semantic similarity is for finding non-obvious relationships across completely different knowledge domains.

What's next for Meridian

PDF and URL ingestion, user authentication with Firebase, belief tracking over time (watching how your positions shift), and an Obsidian vault integration so users can connect their existing notes.

Built With

Share this project:

Updates