Meridian

Meridian is an open-source research workflow for finding professional pain points in a specific industry and turning them into weekly opportunity briefings.

The system is designed around a simple idea: AI is better used to surface signal than to pretend it has answers. Each run is human-steered. You choose a focus, add your own context if you want, and Meridian gathers recent sources, extracts recurring pain points, and synthesizes a briefing you can review in a self-hosted dashboard.

Repository Structure

• config/ industry source packs and focus rotation definitions
• db/ PostgreSQL schema for industries, briefings, and extracted pain points
• n8n/ workflow exports for ingestion, filtering, and LLM synthesis
• dashboard/ a Next.js frontend for browsing weekly briefings

See ARCHITECTURE.md for the current system shape, data boundaries, and planned workflow path.

Current Status

The repository is in a very early phase. Initial work is focused on:

• Defining the industry config contract
• Setting up the database schema
• Replacing mocked workflow steps with live source, LLM, and database integrations

Committed so far:

• A journalism source config
• The initial PostgreSQL schema
• An n8n workflow with a real form-triggered entry step, config loading from disk, live RSS ingestion, relevance scoring, and live pass-1 Gemini extraction
• A minimal Next.js dashboard scaffold in dashboard/

Local Development

To run the dashboard locally:

cd dashboard
npm install
npm run dev

To run n8n locally with the current live pass-1 extraction step:

npx n8n

The workflow now expects the journalism config at config/journalism.json relative to the directory where you start n8n. The intended local flow is to run npx n8n from the repository root.

After importing the workflow:

• submit the built-in n8n form trigger to choose a suggested focus, optional custom focus, and optional context
• create an n8n Header Auth credential with x-goog-api-key as the header name and your Gemini API key as the value
• attach that credential to both Gemini Pass 1 Extraction and Gemini Fallback Extraction

The current workflow also throttles live pass-1 extraction for testing: it allows up to 6 selected source documents per run, scores candidates for focus relevance before extraction, and spaces requests inside n8n to reduce rate-limit failures. It tries the preview Gemini model first and only falls back to a stable flash-lite model for source documents whose preview requests fail. If no recent source clears the relevance threshold, the workflow currently stops before live pass-1 extraction; Gemini and downstream extraction nodes are not invoked for that execution.

In this slice, the workflow loads the journalism config from config/journalism.json relative to the n8n working directory and uses the selected suggested focus as the source-selection anchor. If a custom focus override is provided, Meridian keeps the selected track for source selection and blends the custom focus text into prompt steering and relevance scoring.

Planned Stack

• n8n for orchestration
• PostgreSQL for storage
• Next.js for the dashboard
• Docker Compose for local self-hosting