How it works
Every 3 hours a GitHub Actions job collects news and research papers from 34 sources, removes anything seen before, and updates this site. AI agents pick what is genuinely new, group it into stories and check it against the original articles.
30
sources
7
AI agents
~60–90
LLM calls a day
150
daily call cap
8×
updates a day
$0
running cost
Architecture
Green dot = working today, grey = coming next. Solid arrows run every 3 hours, dashed ones once each morning.
Every 3 hours
- Collect new posts from every source
- Clean links and remove anything seen in the last 30 days
- Scout agents research the important items
- The analyst scores and groups stories
- This site is rebuilt with the latest news and research papers
The 7 AI agents
Built with LangGraph. The order they run in is fixed in code, so LLM usage stays predictable. Every call is traced in Langfuse.
| Agent | Runs | LLM | Job |
|---|---|---|---|
| Scout · Labs & research | every 3 h | Qwen 3.8 27B · Groq | Checks the AI labs and Hugging Face; reads the full article for important items. |
| Scout · Cloud & platforms | every 3 h | Qwen 3.8 27B · Groq | Checks AWS, Azure and Google Cloud for new GenAI products and features. |
| Scout · Community | every 3 h | Qwen 3.8 27B · Groq | Checks Hacker News and Reddit for launches and discussions. |
| Scout · Video, newsletters & blogs | every 3 h | Qwen 3.8 27B · Groq | Checks YouTube (news only, not tutorials), Substack, Medium and newsletter emails. |
| Scout · X | every 3 h | Qwen 3.8 27B · Groq | Checks posts from curated AI researchers and founders. |
| Scout · Web search | every 3 h | Qwen 3.8 27B · Groq | Searches news sites and lab blogs without feeds (YC, Coimbatore too) for fresh GenAI articles. |
| Analyst | every 3 h | gpt-oss-120b · Groq | Scores, categorizes and groups stories; compares with the last 7 days. |
About 8–11 LLM calls per 3-hourly update: roughly 60–90 a day, capped at 150. All on free tiers (Groq and Google Gemini).
Responsible AI
GenAI Daily covers AI that advances people. These rules are enforced in code and in every AI prompt:
- No harmful content. Nothing sexual or explicit, racist, sexist, misogynistic, hateful, anti-religious or violent, and no gossip or fear-mongering.
- Checked three times. A block list at collection, the policy in every AI prompt, and a final code check before anything is published.
- Only reviewed content. Raw web-search results are never shown; they appear only as stories an AI agent reviewed.
- Transparent. AI-written content is labelled, every story links to its sources, and every AI call is logged.
- Grounded. The AI can't invent links, and every number in a story is checked against its source.
- Private. No accounts, tracking, ads or personal data.
Open source
The code, setup guide and full architecture are on GitHub. You can run your own copy for free in about 10 minutes.