AI News Brew

AI News Brew started in 2022 from a simple frustration: the news was everywhere, but clarity was rare. Feeds were optimized for engagement, and the same story came spun six ways depending on who told it.

The idea was to use AI to remove friction, not to replace judgment: pull from dozens of sources at once, find where they agree, flag where they diverge, and distill what actually happened into something you can read in a minute. The haiku was a discipline, not a gimmick. If you can't say it in seventeen syllables, you haven't found the signal yet.

It was my first AI product in public, and it ran daily for three years.

• The prototype (May 2023). A set of Python tools I called Vahanas: pull the last 24 hours of headlines, search ranked terms across many outlets, then have a model walk the coverage of one story source by source and write a single narrative from the best-supported facts. It also scored how similar two outlets' accounts were, which became the seed of the bias reading.
• Aggregation at scale. A news API under a research license fed the pipeline. At its peak it processed thousands of articles a day across dozens of sources.
• Cross-referencing and scoring. Stories were grouped across outlets, summarized with numbered citations back to each source, and given a left-to-right bias reading and a quality score.
• The haiku layer. Every story got a three-line haiku and a generated image, which became the project's signature on the site and on social.
• Agent-assisted editing. By late 2024 a set of CodeGPT agents handled review and publication steps, with me approving what shipped (see the workflow video below).
• The official reader (September 2024 onward). A Next.js web experience, 751 commits, with search, categories and animated reading views. It was the public face of the feed until the AI publishing system stopped.

In late 2025 the research license behind the aggregation ended with no option to renew, and the publication had to close. The site now keeps all 5,738 articles searchable as an archive.

In March 2026 I brought it back smaller: a local pipeline that pulls GDELT and RSS, clusters stories by shared entities, drafts with a local model through Ollama, sets the haiku over an image, and posts to Bluesky through n8n, with a human approving every post.

Taking an AI product from idea to daily operation and keeping it running for three years: sourcing, editorial rules, scoring, publishing, and a reader people could actually use. It also shows what happens when a product depends on someone else's license, and how I responded: preserve the work as an archive, then rebuild the core idea on infrastructure I control.