Tonari Tutor has been in private testing for a few months now, and it’s worth documenting where things actually stand. Not a launch post — we’re pre-alpha v2.0 — but a snapshot of a tool I’ve been using daily to prep for real interviews.

Current state

Four languages: English, Japanese, Spanish, and Mandarin. Two interviewer personas — Kohei and Ai — each with distinct conversational styles. A full feedback loop that scores your responses across multiple dimensions and gives you a structured breakdown after each session.

The feedback scoring turned out to be more useful than I expected. The system generates a “lean hire / lean no hire” recommendation alongside the detailed rubric scores, and after running maybe fifty sessions against it, that binary signal maps surprisingly well to the assessment rubrics real interviewers use. When Tutor says “lean no hire” on my behavioral answers, I know exactly which ones to rework. When it says “lean hire” on a technical walkthrough, the reasoning it gives matches what I’d expect from a senior interviewer’s written feedback.

I’ve been using it to prep for roles at companies where the interview is conducted partly in Japanese. The language switching — described back in the first Tutor post — holds up. The agent stays in character regardless of which language you’re speaking, and the feedback comes back in the session language.

SEO groundwork

On April 2nd I pushed a batch of commits to make Tutor discoverable. The problem with SPAs is that crawlers see an empty div. So I added:

  • JSON-LD structured data describing Tutor as a SoftwareApplication — lets Google understand what the page is without executing JavaScript.
  • Static HTML fallback content for crawlers. The actual app hydrates over it, but bots get real text instead of a blank shell.
  • Sitemap and robots.txt updates. Basic stuff, but I’d been putting it off.
  • Canonical and hreflang tags for the multi-language support.
  • GA4 analytics with custom interview tracking events — session starts, completions, language selection, feedback views.

Six commits in one day. Should have done this months ago. The static fallback approach is the same pattern I used on the widget-to-WebSocket migration — own the HTML, let the framework enhance it.

The next day I added a manual feedback option for when the webhook fails. Edge case, but it matters: if the post-interview feedback webhook drops the response, users can still trigger a retry instead of losing the whole session’s data.

What’s next

Three things on the immediate roadmap:

Professional voice clone. I’m booking an SM57 recording session to capture a proper narrator voice for Tutor’s non-interview audio (instructions, transitions, feedback delivery). The current TTS works, but it doesn’t match the quality bar of the interviewer voices.

More interviewer types. Kohei and Ai cover general and conversational styles, but I want dedicated technical and behavioral personas. Different question patterns, different follow-up strategies, different scoring emphasis.

Possibly opening access. Tutor has been private since day one — invite-only, gated behind an access worker. I’m considering a limited public beta, but only after the voice clone and additional personas are in place. No point opening the door if the experience isn’t where I want it.

You can see the current state at tutor.tonari.ai.