An agent that keeps a LinkedIn presence alive without turning into a content treadmill: it finds what's worth saying, drafts it in my voice, fact-checks it, and waits for my approval in Telegram before anything goes out.
Posting consistently on LinkedIn is mostly a research-and-drafting problem, and automating it naively produces generic, off-voice, or factually shaky posts. The brief was a pipeline that does the legwork — find a genuinely interesting AI topic, draft a few angles that sound like me, fact-check them — but never publishes anything I haven't personally approved.
A Python pipeline pulls live discussions from Hacker News, ranks them for relevance, and drafts three voice-matched angles for the top topic, each with a fact-check pass. Drafts land in a Telegram bot for review: I approve, edit, or reject from my phone, and an approved draft is rendered into a text-card image before it posts to LinkedIn via OAuth.
The human stays in the loop by design — nothing publishes without an explicit approval. The writer learns from what actually ships: published posts and my edits feed a voice corpus, and recorded performance (reactions, comments) feeds back into ranking and drafting. Token expiry is handled gracefully, with the bot warning about LinkedIn re-auth roughly a week out, and forty-six offline tests cover the pipeline end to end.
The result is a posting habit that survives a busy week without sacrificing quality or accuracy. The agent removes the blank-page and research cost, while the approval gate and fact-check keep the standard high and the voice consistently mine — and because it learns from published posts and edits, the drafts get closer to first-pass-ready over time.
The full architecture, the voice-learning approach, and a walkthrough of the research-to-publish cycle can be shared for review.