Telegram Intelligence System

Open-source intelligence pipeline for Telegram, X and Obsidian. It monitors Telegram sources available to an authorized user account, filters and deduplicates messages, builds daily digests, analyzes X links, enriches X metrics, discovers professional profiles, writes structured Obsidian notes and exposes a small dashboard.

Components

• apps/telegram-offer-parser — Telegram monitoring, vacancy/knowledge filters, semantic deduplication, Twitter radar and daily digest.
• apps/obsidian-bot — Telegram inbox bot that structures manually submitted notes and writes Markdown to an Obsidian vault.
• apps/miniapp — read-only web dashboard for recent results and metrics.
• services/people-radar — opt-in professional discovery from configured chats.
• services/x-collector — read-only browser collector for public X post text and engagement metrics.

Safety and platform limits

• Use only accounts and chats you are authorized to access.
• Respect Telegram flood waits, X automation rules and local privacy law.
• The repository intentionally contains no sessions, cookies, tokens, chat exports, databases, vault contents or collected people data.
• The X collector blocks known write actions, but operators remain responsible for account security and platform compliance.

Quick start

Requirements: Linux, Python 3.11+, systemd (optional), a Telegram api_id and api_hash.

git clone https://github.com/Floopi10/telegram-intelligence-system.git
cd telegram-intelligence-system
cp .env.example .env
python3 -m venv .venv
. .venv/bin/activate
pip install -e apps/telegram-offer-parser
offer-parser --help

Fill .env locally. Never commit it. Component-specific instructions are kept in each component directory.

Quality-first digest

Twitter Radar does not fill a fixed quota. It separates metric-confirmed winners, early signals and promising mechanics, requires visible and internally consistent X metrics, ranks by velocity and engagement before LLM score, and drops low-view posts that only received a high text-analysis score. A post needs at least 500 views to become a relative winner; posts with 100-499 views can only be labelled as early signals. Vacancy and knowledge sections also re-check geography, role relevance, campaign spam, greetings and conversational fragments before the daily digest is rendered.

When a Telegram message explicitly reports an outcome next to an X link, the radar also extracts follower gain, the result window, views-to-followers conversion, revenue and content format. These claims remain labelled as self-reported: they improve ordering and explainability, but cannot make a post a winner without metrics collected directly from X.

Each digest covers everything since the previous one (at most 72 hours) and re-reads X metrics for up to 30 candidates first, so a post is judged by its current result rather than the moment it was first seen. A later snapshot can confirm a post but never demote it.

LLM analysis of X posts is limited to protect the provider's daily quota: engagement requests and a third post from one author per UTC day are skipped, and TWITTER_AI_DAILY_BUDGET (default 680) is spread evenly across the day. Priority chats skip the filters and the pacing. A skipped URL can be analysed on its next mention.

Digest delivery retries temporary Telegram failures every five minutes instead of silently stopping, and a retry skips the parts already delivered. Instant notifications go through a persistent outbox, so a Telegram failure or flood wait delays them instead of losing them. Operators can perform a controlled catch-up with offer-parser send-digest --force while the monitor service is stopped.

License

AGPL-3.0. Network-hosted modified versions must make their corresponding source available under the same license.