How Voxline is built
The architecture, the guardrail decisions, and what's changed release by release — for anyone who wants more than the demo.
Architecture
Next.js and Vercel for the app, Supabase for auth and data, and the Claude API for both the voice analysis and the drafting itself. The one decision worth explaining is where content moderation sits.
You paste a comment you want to reply to.
Claude drafts a reply in your voice profile.
The draft — not the comment — is checked for tone.
You edit or approve. Nothing posts on its own.
Moderation runs on the drafted reply, not the pasted comment. Early versions gated on the comment's tone, which meant a hostile comment could block someone from replying to it at all — punishing the person being harassed, not the one doing it. Checking the draft instead fixed that.
Release notes
Opening up access
- Removed the invite-code requirement — signups are now open, capped at 100 users total instead of gated behind a code.
- Added Cloudflare Turnstile to the signup form to stop scripted signups before they ever reach account creation.
- Moved from a shared subdomain to voxline.app, with its own support email.
Chrome extension
- Shipped a Chrome extension — paste a comment, get a draft, copy it out, without leaving the page.
- Drafting can now speak as the post's owner or as a peer commenter, depending on whose post you're replying to.
Real evals, not vibes
- Built a scoring pipeline that grades every draft on voice fidelity, factual accuracy, and template diversity, with a hard fail for anything that invents facts.
- Caught and fixed two real defects the eval surfaced: drafts quoting a single writing sample's exact phrasing as if it were general vocabulary, and drafts speaking as if the user owned a community they were only commenting in.
- Shortened data retention from 30 days to 7, with feedback (thumbs up or down) kept longer than the comment text it was about.
Privacy hardening
- Stopped storing raw writing samples after voice analysis — only the derived score, traits, and style guide are kept.
- Added a plain-language privacy page, a “Delete my data” button, and a daily job that clears anything older than 30 days.
- Added a lightweight scrubber that redacts emails, phone numbers, and card-like numbers before anything is saved.
Guardrail redesign
- Built a 58-case test set to stress-test content moderation — sarcasm, backhanded compliments, and the hardest case: someone reporting hostility said to them.
- Compared three moderation approaches and fixed a real false positive with a two-signal design (toxicity, and whether it targets a person).
- Redesigned enforcement to check Voxline's own drafted reply instead of the incoming comment, after realizing the original design punished the person being harassed, not the one doing it.
Rebuilt on a self-owned stack
- Migrated off a no-code platform onto Next.js, Supabase, and the Claude API, with full control over auth, data, and deployment.
- Shipped the core loop: paste a writing sample, get a voice score, draft replies in that voice.
Voxline is capped at 100 users while it grows. Sign up.