
TEAHOUSE News Bot
Telegram bot and web admin for the TEAHOUSE fintech channel: it collects the week’s news from nine sources, has AI score, sort and write the digest, makes a cover and posts to the channel on schedule.
Challenge
The TEAHOUSE community runs TH NEWS POST, a Telegram channel with a weekly fintech digest. Every issue means reading a dozen news sites, picking the stories that matter, writing a dozen headlines in the channel's format, finding a cover and posting on time — every single week. The brief was to automate that routine without handing the channel over to a model: the editor keeps the last word, the AI must never invent a story or a link, and every post has to fit Telegram's formatting and length limits.
Solution
- News harvested from nine sources — Finextra, The Paypers, Finance Magnates, CoinDesk, Latvijas Banka and others — via RSS, sitemaps or page parsing; paste any URL and the bot detects the feed type itself
- Duplicates dropped by content hash, old articles pruned on a schedule, and every source weighted so stronger outlets rank higher
- AI editor in three passes: it puts each article into a category, scores it 0–100 with a one-line reason against editorial criteria the client writes in plain words, then writes the digest from the best-scored stories
- Hallucination guard: the model refers to stories by id, never by URL — the server substitutes the real links and drops any line whose link it cannot verify
- Provider-agnostic text AI — Groq (Llama, GPT-OSS), Anthropic, OpenAI or Google Gemini — switched in settings, with a connection test
- Covers from a photo library sorted by category, or generated by AI through Pollinations (Flux), Cloudflare Workers AI, Google Imagen or fal.ai
- Every Sunday the bot collects the week and builds a draft; the editor gets a preview in Telegram and publishes with one tap — now, in the default Monday-morning slot or at any time — with an optional heads-up before it goes out
- "Tea posts": standalone posts the AI writes in the channel’s own voice from a one-line topic, with a generated cover
- Telegram-native output: MarkdownV2 escaped automatically, captions and messages kept inside the 1024 / 4096-character limits, link previews instead of bare URLs
- Web admin with a health dashboard, pipeline and calendar, articles with their AI scores, sources, photo library, channel-reaction analytics and an operator guide in Russian
- Optional cross-posting of each digest to LinkedIn
- Self-hosted in Docker on a VPS behind Caddy with automatic TLS, push-to-deploy and daily database backups
Result
Putting the weekly digest together now comes down to reviewing a draft and one tap in Telegram: on Sunday the bot collects and ranks the week's fintech news, and by Monday morning the issue is on the channel. Editorial control stays with a person — the criteria are plain text the client edits, every score comes with its reason, and the model cannot put an invented story or link on the channel.
More apps work
Related case studies
Want a similar project?
Let's discuss how we can bring your vision to life with the same quality and attention to detail.
Get in Touch


