trend-radar

Pull signals from public sources where trends surface first: the Algolia HackerNews API, HuggingFace trending papers and models, and curated newsletters via Gmail. Normalize items into canonical topics, score with recency decay and cross-source bonuses, and rank the results.

Query hn.algolia.com with tags=front_page for current momentum, or search_by_date with a unix timestamp cutoff and a points floor for recent stories. Compute a velocity proxy as points divided by age in hours to rank fast-rising stories.

Each item gets a score of source weight times recency decay (7-day half-life) times normalized velocity. Topic scores sum their items and multiply by a bonus of 1.0, 1.5, or 2.0 depending on whether one, two, or three sources corroborate the topic.

Gmail's API does not match nested labels by their leaf name, so label:newsletters fails while label:Content/newsletters works. If the full path still returns nothing, fall back to querying the curated newsletter sender addresses directly.

No. The workflow is human-in-the-loop by design: it presents a ranked digest and waits for explicit approval before searching for duplicates and writing trend notes to the Notion Topics database. Net-new topics require separate confirmation.

It cannot read LinkedIn-native trend or engagement data because no legitimate API exists and scraping is ruled out. The proxy sources (HN, HuggingFace, newsletters) cover AI and tech momentum well but are not a substitute for LinkedIn parity.