SiMon

Personal geopolitics dashboard. Reads Telegram channels, summarizes events with an LLM, and shows them on a web dashboard organized by country, as a self-updating "since last read" catch-up plus a per-week archive.

Architecture

Three pieces share a database so nothing has to run on your machine:

• Vercel serves the dashboard and the read/write API (app.py, entry api/index.py).
• GitHub Actions runs the worker (worker/refresh.py) on a schedule and on demand: it fetches Telegram messages and summarizes them.
• Turso (libSQL) is the shared database both sides use.

How it works

1. The worker connects to Telegram as a regular user account (not a bot) via the Telethon User API and fetches messages since the last run.
2. New messages are summarized with DeepSeek-V3, merged incrementally into the report (only the delta plus the existing summary is sent, keeping cost down) with inline source citations.
3. Each run updates two reports: the current week's report (one per UTC-Sunday week, kept as an archive) and the rolling "since last read" catch-up.
4. The dashboard shows the catch-up on top and the weekly archive below, with per-country tabs. Citations link back to the news outlet or the Telegram channel.

Repo layout

Path	Purpose
app.py, api/, public/	Web layer: API + dashboard, served by Vercel
core/	Shared modules (db, weeks) used by both runtimes
worker/	Telegram fetch + summarization, run by GitHub Actions
scripts/gen_session.py	One-time: generate the Telegram session string
tests/	Unit tests (pytest)

Deploying

See DEPLOY.md for the full Turso + secrets + Vercel setup.

Local development

python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt -r requirements-worker.txt

Create .env:

TELEGRAM_API_ID=your_api_id
TELEGRAM_API_HASH=your_api_hash
DEEPSEEK_API_KEY=your_key

With no TURSO_DATABASE_URL set, the app uses a local dashboard.db file. Authenticate Telegram once (python -m worker.telegram_fetcher saves a session file, or python scripts/gen_session.py prints a portable session string), then:

python -m worker.refresh   # fetch + summarize into the database
python app.py              # serve the dashboard at localhost:8000

Run the tests with pytest.

Cost

DeepSeek-V3 pricing: $0.28/MTok input, $0.42/MTok output. Incremental merging keeps each refresh small; a typical run costs well under $0.01. Per-report cost is shown in the dashboard header.