the-resume-thing

The problem

Solo and open-source developers have no manager, no performance review and no employer brand vouching for them, and their work is scattered across half a dozen tools. The difficulty isn't writing a résumé — it's legibility. Proving what you did.

Meanwhile the obvious fix, "ask an LLM to write my résumé", produces confident sentences nobody can check, which is worse than nothing in the one setting where you'll be asked to defend every line.

The approach

It connects to the places you actually work — GitHub repos, PRs, issues and commits; Linear issues and projects — and normalises them into one provider-agnostic activity graph. An LLM drafts accomplishments from that graph, and every bullet carries the PR, issue or commit it came from. The model is constrained to the evidence rather than trusted to be honest.

You then review, rather than accept: suggestions arrive as a keyboard-driven queue you apply, edit, regenerate or dismiss.

The output is a real typeset document, not an HTML approximation — LaTeX rendered to PDF through Tectonic, with the source editable in a built-in CodeMirror editor with live preview.

The hard part

Making the model unable to invent things. "Don't hallucinate" is not an implementation. The design does the work instead: claims are synthesised only from ingested activity, and each one keeps a link back to its source, so an unsourced bullet can't exist by construction. Verifiability is the feature — everything else is convenience.

Syncing without destroying your edits. A background worker keeps pulling new work, which means the interesting case isn't the first run, it's the hundredth. Drafting is additive: new suggestions accumulate alongside what you've already accepted and edited, and never overwrite it. Get that wrong once and the tool eats an afternoon of someone's writing.

Where it stands

Early and evolving, built in the open. Next.js 16 and React 19 on Prisma and SQLite, with GitHub and Linear OAuth wired through Better Auth.