AI-assisted News Briefing

The problem

News about AI and automation is high volume. Reading everything manually produces a feed with little connection to practical product decisions, portfolio work, automation ideas, or professional development.

The goal was a small briefing: fewer items, better relevance, and a clear reason each item matters.

Key decisions

Rank before generating

Collection, normalization, history filtering, deduplication, and relevance ranking are explicit code paths. A language model cannot reintroduce an item that the deterministic rules excluded.

Make the model optional

The pipeline can generate a basic briefing without a Gemini key. When configured, Gemini improves summaries rather than owning source selection or delivery safety.

Preview before delivery

The default workflow creates a local preview. Telegram delivery requires an explicit send command and remains separate from a scheduled job.

Select fewer, stronger items

The pipeline selects up to five useful stories instead of filling a fixed quota with weak matches.

Solution flow

Public RSS, Atom, and supported HTML sources
-> collection and normalization
-> duplicate and delivery-history filtering
-> deterministic relevance ranking
-> optional Gemini summary generation
-> Portuguese Telegram-ready digest

Engineering and reliability

• Node.js and JavaScript.
• RSS, Atom, and HTML collection.
• Local delivery history outside Git.
• Optional Gemini and Telegram integrations.
• Manual GitHub Actions entry point. An earlier automatic schedule was switched off in August 2026 until a supervised delivery round validates the output.
• Syntax validation for JavaScript entry points.

Outcome and evidence

• The repository documents a reproducible local preview path.
• The model-free fallback keeps the core workflow usable without credentials.
• The implementation records sent items locally to reduce duplicates.
• Automation remains intentionally supervised rather than being described as a fully autonomous production system.

What I learned

A useful AI-assisted workflow needs explicit quality gates before generation. The model is more valuable when it works inside a narrow, observable system than when it is asked to decide every part of the process.

What I would improve next

• Run a supervised delivery round and evaluate the digest quality with explicit criteria.
• Measure useful feedback only after deciding what should count as a meaningful signal.
• Add new sources only when their public endpoints and relevance fit the existing validation rules.