JobMe: AI Mock Interview Platform

JobMe is a 3D mock interview office: an AI recruiter reads your résumé, hears your answers, follows up on the gaps, and scores you live, so your real internship interview isn't your first.

Comment

Inspiration

Behavioral interviews usually go wrong on the follow-up, not the first question: "You said the team fixed it. What did **you* change?"* Most practice tools never ask that. Question lists and flashcards don't push back. A chatbot told to "act as an interviewer" drifts off and doesn't score you. A mock interview with a real person means booking a career-center slot or finding a friend who has done one. Students who don't have that access, including first-generation students and those at schools with small career centers, almost never get to practice the follow-up.

We wanted a room where the recruiter actually listens. A vague answer and a sharp answer to the same question should get different next questions, and you should be able to see why.

What it does

JobMe is a mock interview for behavioral internship practice, set in a small 3D office you walk through.

  • Set up at the front desk. Choose Quick or Full. You can add a PDF résumé and a target job, either pasted or imported from a Greenhouse, Lever, Ashby or SmartRecruiters link. Questions come from a 15-topic bank and from your own projects and roles.
  • Answer out loud or by typing. The recruiter speaks each question with captions and lip-sync. You answer by voice or by typing.
  • Watch the notepad. After each answer, the recruiter's notepad shows your scores, the main gap, the quote that shows it, and why the next question was chosen.
  • Get the follow-up you need. "We worked as a team and fixed the bug" gets a question about your own part. A specific answer with a clear result gets a deeper question instead.
  • Review and come back. Each interview produces a report with overall and per-question scores, what went well, what to work on, and a STAR rewrite of your weakest answer. A ledger keeps every transcript word for word, and trends show whether your weak spots are improving.

How we built it

  • Next.js 16, React 19 and TypeScript for the UI and API routes in one codebase.
  • A deterministic interview engine. The AI scores answers and writes the wording, but fixed rules choose the next move: clarify, go deeper, switch topic, or wrap up. This keeps the adaptivity predictable and testable, and the notepad's explanation always matches what actually happened.
  • Google Gemini scores each answer against a rubric that depends on the question type (behavioral, technical or motivation). The same call drafts both possible follow-ups, so each turn needs only one request. Gemini also plans personalized opening questions, reads résumés and job descriptions, writes the report text, and embeds résumé chunks.
  • Supabase handles auth and Postgres, with row-level security. A submit_turn database function saves each answer and the engine state together in one transaction. Sequence checks mean a double-click or a flaky network never saves a turn twice. Résumé embeddings are stored in pgvector.
  • Three.js, React Three Fiber, Drei and GSAP build the office: a street with a time-of-day sky, a front desk with a receptionist (Ms. Crystal), and an interview room with a toon-shaded recruiter. Her mouth follows her words using character timings from ElevenLabs.
  • Privacy by default. No audio is stored. Résumé PDFs are parsed, redacted and then discarded.
  • Vitest tests cover the engine, turn and report handling, résumé helpers and voice logic. Offline rehearsal scripts and live verification scripts are also included.

Challenges we ran into

  • Latency on a free tier. A live interview has to feel live, but Gemini's free tier allows only 15 requests a minute. We wrote a latency probe to choose the model, logged timings for every turn, and gave scoring 8 seconds. If a score is late, the interview moves on and the score is attached to its answer when it arrives. Any answer still unscored is scored again when the report is filed.
  • Scoring spoken answers fairly. An audit found spoken answers were rated weak because of their format: Chrome's transcripts have no capitals or punctuation, STAR order was being scored, and answers without numbers lost points. We replaced the STAR-only rubric with one per question type that never penalizes grammar, accent, filler words or a missing number.
  • Planning for failure. Any outside service can fail in the middle of an interview, so each has a fallback. Fixed rules keep the conversation going when scoring fails. Browser speech replaces ElevenLabs. Typing replaces speech recognition. The 3D office has a 2D version.
  • Making a 3D office feel calm, not gimmicky. Camera walks, lip-sync, a page-turning ledger and a receptionist who greets you once per sign-in took many rounds of fixing race conditions, flicker and focus bugs.

What we learned

  • Let the model do what it's good at, judging and wording, and keep the control flow in code. That made the product explainable and much easier to test.
  • Measure before choosing. A latency probe, not a guess, picked the model.
  • Rubrics carry bias. Scoring transcribed speech like an essay was quietly penalizing people for talking.
  • Plan for failure from day one. A demo that keeps working when something breaks beats one that only works when everything goes right.

What's next

Recruiter personas (a warm campus recruiter or a tough senior engineer, on the same engine), hands-free turn-taking with voice-activity detection, a shareable progress report for a mentor or career center, and a spoken technical track such as "explain your code."

Built With

Share this project:

Updates

Submission history