title	BlogAI
emoji	✍️
colorFrom	green
colorTo	blue
sdk	gradio
sdk_version	6.10.0
app_file	app.py
pinned	false
short_description	Turn an article + your notes into a blog post draft

BlogAI

You read something interesting. You have thoughts. BlogAI turns that combination into a draft blog post and LinkedIn post — in your voice, ready to edit and publish.

Built with Gradio + Qwen 2.5 7B via HuggingFace Inference API.

Live on HuggingFace Spaces

---

What this is for

You already read articles and have opinions about them. BlogAI removes the gap between "I should write about this" and an actual draft. The use case is personal brand building — posting consistently without the blank page problem.

The core loop:

1. Read something on your phone → send the URL + your reaction to the Telegram bot
2. Later, open BlogAI → paste the URL and your notes → generate a blog post and LinkedIn draft
3. Edit, approve, post

---

What this is NOT for

• Writing from scratch — BlogAI needs an article URL and your notes as input. It is not a general writing assistant.
• Paywalled or JS-heavy articles — NYT, Bloomberg, most major newspapers require subscriptions. The app will fail to fetch those. Use open-access sources (Wikipedia, dev blogs, newsletters, arXiv, company announcement pages).
• Publishing directly — BlogAI generates drafts. It does not post to Medium, LinkedIn, or anywhere else. Copy, edit, post manually.
• Replacing your voice — the draft is a starting point. The model (Qwen 7B) will follow your notes but may drift from your exact style. Always edit before posting.

---

How to use

Capture (mobile — recommended)

Send a single message to @tayapb_bot on Telegram:

https://example.com/article-you-read

Your raw reaction here. As rough as you want.
Can be multiple lines.

The bot extracts the first URL it finds and treats the rest as your notes. It saves both to your Notion drafts queue with status Inbox.

If you have no URL yet — notes only is fine too. The bot will save them without a link.

Generate (web app — HuggingFace)

1. Open the BlogAI Space
2. Paste the article URL and your notes
3. Pick a tone (Blog / Social is the default)
4. Click Generate draft

Run locally

git clone https://github.com/tyakovenko/blogAI.git
cd blogAI
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
export HF_TOKEN=your_hf_token
export ANTHROPIC_API_KEY=your_anthropic_key  # optional fallback
python app.py

Opens at http://localhost:7860. No GRADIO_SHARE needed — the tunnel is only enabled on HF Spaces.

You get two outputs:

• Blog Post — 300–500 words, personal voice, plain paragraphs
• LinkedIn — 150–250 words, hook opening, single closing question

The LinkedIn-ify button (experimental) is a separate tool — it sends your raw notes to Kagi Translate for a different style rewrite. Not the same as the LinkedIn draft above.

Tone options

Tone	Best for
Blog / Social	Personal posts, casual takes, dev blog
Professional	Work-adjacent writing, industry commentary
Academic	Structured analysis, course work adjacent

---

Known limitations

• Telegram requires the HF Space to be awake. The bot runs as a background thread inside the HF Space process. If the Space goes idle (~15 min of no web traffic on free tier), the thread dies and incoming messages are silently dropped — no error, no notification, no retry. To wake it: visit the Space URL before sending messages. For reliable always-on capture: (a) pipe Telegram → Notion via Make.com or Zapier (free tier, no code); (b) migrate the bot to a webhook deploy on Render or Fly.io independent of the Space.
• Paywalled and JS-rendered pages will fail — the app returns an error. Open-access sources (dev blogs, newsletters, arXiv, company pages) work fine.
• Generation model quality varies — HF free tier models follow formatting rules imperfectly. Falls back to Claude Haiku automatically if HF Inference is unavailable (requires ANTHROPIC_API_KEY set in Space secrets).
• The app generates drafts only — it does not post to Medium, LinkedIn, or anywhere else.

---

Evaluation

The Qwen→Haiku pipeline was evaluated on 27 samples across two output modes (blog, LinkedIn) against Haiku standalone and Qwen standalone baselines. The pipeline outperforms Haiku standalone on substance fidelity (significantly on LinkedIn: 0.663 vs. 0.624, p=0.030) and substantially on voice fidelity (Tier 2 structural score: 0.956 vs. 0.800). The LinkedIn edit pass costs 38% less than standalone Haiku generation.

Full methodology, results, and data: blogAI_evals