Inhuman Science

Telegram | Twitter/X

Automated AI/ML content curation and publishing pipeline. Aggregates papers, blog posts, and tweets from the AI world, evaluates them with LLMs, and publishes the best finds to Telegram and Twitter/X.

How It Works

The system runs four scheduled jobs (cron-based, timezone-aware):

Papers (10:00) — scrapes trending papers from AlphaXiv, scores them with an LLM oracle, downloads PDFs, extracts the most representative figure using a vision model, generates bilingual posts (Russian for Telegram, English for Twitter), and publishes.

Blogs (12:00) — fetches RSS feeds from 11 sources (OpenAI, Anthropic, Google Gemini, Google DeepMind, Meta AI, Microsoft Research, NVIDIA Tech, Amazon Science, IBM Research, Apple ML, PyTorch), scores and fact-checks each post, generates summaries, and publishes.

Twitter (14:00) — monitors 25 accounts: AI lab leaders (Demis Hassabis, Jeff Dean, Dario Amodei, etc.), top researchers (Yoshua Bengio, Fei-Fei Li, Noam Brown, Jan Leike), open-source leads (Thomas Wolf, Clément Delangue, Soumith Chintala), and official accounts (OpenAI, DeepMind, Anthropic, Meta AI, Mistral, Cohere, ByteDance OSS). Scores tweets, generates Russian summaries for Telegram, retweets on Twitter.

Backup (03:00) — daily SQLite dump to backups/ with Telegram delivery. Keeps last 7 days.

Content Filtering

The oracle is tuned for pure science and engineering only. It publishes new models, architectures, agents, benchmarks, open-source releases, and infrastructure breakthroughs. It rejects politics, business deals, gossip, marketing, and think-pieces.

Cross-Source Deduplication

When the same news goes viral across multiple sources, a batch dedup step groups all scored candidates by topic in a single LLM call and keeps only the highest-scored item per group. This runs against both the current batch and the last 5 days of published content.

Architecture

Sources                 Processing              Publishing
┌─────────────┐        ┌──────────────┐        ┌───────────┐
│  AlphaXiv   │───┐    │  Oracle      │        │ Telegram  │
│  (papers)   │   │    │  (scoring,   │   ┌───▶│ (RU post) │
├─────────────┤   │    │  fact-check, │   │    ├───────────┤
│  RSS Feeds  │───┼───▶│  batch dedup)│───┤    │ Twitter   │
│  (11 blogs) │   │    ├──────────────┤   └───▶│ (EN post) │
├─────────────┤   │    │  Processors  │        └───────────┘
│  Twitter    │───┘    │  (PDF, image,│              │
│  (25 accts) │        │  post gen)   │              ▼
└─────────────┘        └──────────────┘        ┌───────────┐
                                               │  SQLite   │
                                               │  (state)  │
                                               └───────────┘

Project Structure

├── main.py                 # Entry point, scheduler, pipeline orchestration
├── config.py               # Configuration and environment variables
├── Makefile                # Convenience commands (deploy, logs, backup, etc.)
├── requirements.txt
├── Dockerfile
├── docker-compose.yml
│
├── sources/
│   ├── base.py             # ContentItem dataclass
│   ├── alphaxiv.py         # AlphaXiv trending papers scraper
│   ├── blogs.py            # RSS feed parser
│   └── twitter_feed.py     # Twitter API v2 feed reader
│
├── oracle/
│   └── oracle.py           # LLM scoring, fact-checking, batch deduplication
│
├── processors/
│   ├── pdf.py              # PDF download and text extraction
│   ├── images.py           # Best figure extraction via vision model
│   └── post_generator.py   # Bilingual post generation (RU/EN)
│
├── publishers/
│   ├── telegram.py         # Telegram channel publisher
│   └── twitter.py          # Twitter/X publisher
│
├── storage/
│   └── state.py            # SQLite state tracking
│
└── llm/
    └── client.py           # OpenRouter API client

Setup

Prerequisites

• Python 3.11+
• API keys: OpenRouter, Telegram Bot, Twitter/X

Configuration

cp .env.example .env
# Fill in API keys

Required variables

Variable	Description
OPENROUTER_API_KEY	OpenRouter API key for LLM calls
TELEGRAM_BOT_TOKEN	Telegram bot token
TELEGRAM_CHANNEL_ID	Target Telegram channel ID
TELEGRAM_ERROR_CHAT_ID	Chat ID for error notifications
TWITTER_API_KEY	Twitter API key
TWITTER_API_SECRET	Twitter API secret
TWITTER_ACCESS_TOKEN	Twitter access token
TWITTER_ACCESS_SECRET	Twitter access secret

Optional variables

Variable	Default	Description
SCHEDULE_PAPERS_CRON	0 10 * * *	Papers pipeline schedule
SCHEDULE_BLOGS_CRON	0 12 * * *	Blogs pipeline schedule
SCHEDULE_TWITTER_CRON	0 14 * * *	Twitter pipeline schedule
SCHEDULE_BACKUP_CRON	0 3 * * *	DB backup schedule
TWITTER_MONITOR_USERS	25 accounts	Comma-separated Twitter usernames
ORACLE_MIN_SCORE	7	Minimum LLM score (1-10) to publish
ORACLE_MAX_PAPERS_PER_RUN	5	Max papers published per run
ORACLE_MAX_BLOGS_PER_RUN	3	Max blog posts published per run
BACKUP_DIR	backups	Directory for DB backups
BACKUP_KEEP_DAYS	7	Days to retain backups
TIMEZONE	Europe/Moscow	Timezone for scheduling

Usage

Docker (recommended)

make deploy     # Build and start
make logs       # Follow logs
make stop       # Stop
make restart    # Restart
make update     # git pull + rebuild
make backup     # Manual DB backup
make status     # Container status

Manual

pip install -r requirements.txt

python main.py              # Start scheduler (runs forever)
python main.py papers       # Single papers run
python main.py blogs        # Single blogs run
python main.py twitter      # Single twitter run
python main.py backup       # Manual backup
python main.py all          # All pipelines sequentially

LLM Models

All LLM calls go through OpenRouter. Models are configured in config.py:

Task	Model
Content scoring	deepseek/deepseek-chat-v3-0324
Fact-checking	deepseek/deepseek-chat-v3-0324
Post generation (RU & EN)	anthropic/claude-sonnet-4.6
Figure extraction (vision)	google/gemini-2.5-flash

Data Storage

SQLite database (state.db) with five tables:

• posted_papers — published papers (arxiv ID, title, timestamp)
• posted_blogs — published blog posts (URL, title, timestamp)
• posted_tweets — published tweets (tweet URL, author, timestamp)
• oracle_decisions — all scoring decisions with scores and reasoning
• published_summaries — title + summary of all published content for cross-source dedup