For anyone preparing for a real interview
Practice the interview you're actually about to have.
Paste the job posting. Lebah checks your CV against it, writes your cover letter, then interviews you out loud — and scores every answer against the requirements the role was written around.
- 2 free credits when you join
- English or Bahasa Indonesia
- Your report is private by default
One job, six steps
- Your profile
- The job
- CV check
- Cover letter
- Interview
- Your report
Demo video coming soon
See it first
Watch a full run before you sign up.
One job posting, start to finish: the CV check, the letter, the interview itself and the report it produces. No account needed to watch.
What you'll see
- 01Your profile — Save your CV and details once — every step below reuses them.
- 02The job — Paste a link to the posting and Lebah pulls the description out of the page.
- 03CV check — See how your CV matches this role, and which gaps actually matter.
- 04Cover letter — A letter built from your real experience, ready to export.
- 05Interview — A 10–15 minute voice interview, with follow-ups that react to your answers.
- 06Your report — A score per requirement, quoted from what you actually said.
How it works
Four steps, one job at a time.
Lebah is built around a single role you're actually applying for — not a library of generic questions to memorise.
Add your profile and the job
Upload your CV once and it is saved to your profile. Then paste a link to the job posting, or the text of it, and Lebah reads out the description and requirements.
You get: a saved profile every other step reuses, and a job description you can edit before anything is generated.
Check the application you're about to send
Lebah compares your CV against that specific posting, and can write a cover letter from what your CV actually says — with a list of the facts it leaned on so you can check them.
You get: a fit score, gaps ranked by how much they matter, and a letter you can export as PDF, Word, text or Markdown.
Sit the interview out loud
An AI interviewer joins you in a video call, asks its questions by voice and listens to your answers. Answer thinly and it digs deeper; answer well and it moves on. Ten to fifteen minutes.
You get: a real spoken interview, recorded, with a full transcript.
Read the report, then go again
Every requirement is scored from 0 to 5 and backed by a quote from what you said, with a note on what a strong answer would have contained. Practise the same role again and the questions are new.
You get: a scorecard you can act on, and a score trend once you have practised more than once.
03Step one
Your JD becomes a structured interview, not a prompt.
An offline planning pass turns the posting into competencies, weights, time budgets, seed questions, follow-up ladders and a scoring rubric — one plan, generated before anyone says hello.
Most AI interviewers improvise. Lebah separates thinking from talking: the hard reasoning happens up front, in a plan the live agent then executes. That plan is validated, not vibes-checked — weights must sum to 1.0, time budgets must add up to the session length, and every requirement carries its own strong/acceptable/weak rubric.
- 3–10 competencies per interview, each traced back to the JD
- Every competency ships with a seed question and a 3-level follow-up ladder
- Attach your CV and Lebah cross-references it — JD requirement, CV overlap, or CV gap: the thing you'll actually get grilled on
- 10–15 minute default session, budgeted per competency
…own high-throughput payment services.
5+ years PostgreSQL, indexing, query plans.
You design APIs and reason about trade-offs.
Distributed caching experience required.
You communicate decisions clearly.
position_summary: Senior Backend Engineer — payments platform, high-throughput services.
04Step two
It talks. You talk back. Camera on.
A realtime voice agent joins you over WebRTC — speech in, speech out, with your camera running the way a real screen would be.
This is the part that makes it uncomfortable in the useful way. You answer out loud, in real time, with no edit buffer and no time to Google. The agent listens, decides how deep to go, and moves on when a competency is covered.
- Sub-second cascaded pipeline: speech-to-text → LLM → speech
- Camera + mic, in-browser, nothing to install
- The whole session is recorded and transcribed as you go
04:12AIWalk me through how you'd scale this payment flow.
04:31YouI'd shard by merchant id and put hot paths behind Redis…
05:02AIWhat breaks first under 10× load?
05The part a question list can't do
Every answer decides the next question.
Each competency has a three-rung follow-up ladder. Answer shallowly and Lebah climbs it. Answer well and it moves on.
A static question list rewards preparation theatre. A ladder rewards depth. Lebah tracks coverage against the plan in real time — which requirements have real evidence, which are still thin, how much time is left — and spends the remaining budget where the interview is weakest.
- 3 escalating probe levels per competency
- Live coverage tracking against the plan's time budget
- Probing style shifts with interview type — STAR completeness for behavioural, technical depth for technical
06Step three
Scored against the JD. Quoted from the transcript.
A post-session pass over the full transcript produces a 0–5 score per requirement, each one backed by verbatim evidence from what you said.
No unattributable “you seemed underprepared.” Every score points at the sentence that earned it, so you can argue with it — which is exactly how you learn what a real interviewer heard.
- Overall score plus a per-competency 0–5 breakdown
- Verbatim quote evidence, strengths and gaps for each requirement
- Delivery patterns: pacing, clarity, filler words
- Sessions that ended early are flagged incomplete, not scored as if finished
“I'd shard by merchant id, then put hot paths behind Redis…”
“…a composite index on (merchant_id, created_at) cut p95 from 800ms to 90ms.”
“We used cache, but I haven't thought about invalidation much.”
“Let me walk you through the trade-off before I pick…”
07Not just a verdict
Told what to fix, per requirement.
Alongside every score, a specific coaching note: what a strong answer to that question would have contained.
A number tells you where you stand. The coaching layer tells you what to do before Tuesday. It's written against the same rubric the planner wrote before the interview started, so “strong” means something fixed rather than whatever the model felt like that day.
Quantifies blast radius and hit rate, names the invalidation strategy, owns the failure mode.
- ✕No cache-invalidation strategy
- ✕Didn't quantify hit ratio or staleness budget
A strong answer names the invalidation policy (TTL + event-based purge), states a hit-rate target, and describes what happens when the cache is wrong — then measures it.
08The long game
Watch a weak competency stop being weak.
Attempts group under the plan they belong to. Lebah tracks the delta per requirement between attempts, and the same competency across different roles.
One interview is a snapshot. Five is a trend line — and the trend is what tells you whether “system design” is a bad day or a real gap. Comparisons run against your previous evaluated attempt, so an abandoned room never pollutes the line.
The rest of the workflow
The interview isn't the only thing standing between you and the job.
The same profile and the same job posting feed three more steps. Nothing has to be pasted twice.
The problem — retyping a job posting into yet another box
Import the job from a link.
Paste the URL of the posting. Lebah reads the page on its own, prefers the structured job data most boards publish, and drops the description into a box you can still edit. No AI guessing, nothing stored — not the page, not the link.
Sites that require a login — LinkedIn and Indeed among them — usually can't be read this way, and Lebah says so plainly instead of returning a page of navigation text. Paste the description instead and everything else works the same.
The problem — not knowing why you never hear back
Check your CV against this role.
Your CV and the posting, side by side: what genuinely matches and where the evidence for it sits, what's missing and how badly, and the words the posting uses that your CV never does.
Fixes are ordered by what to do first, so a CV rewrite is one afternoon rather than an open-ended project.
The problem — a blank page, an hour before the deadline
Write the cover letter from what's true.
From your CV and the posting, or from a rough draft of your own that Lebah tidies up without changing the facts. It arrives with the things it leaned on and the things it deliberately did not claim.
Export as PDF, Word, plain text or Markdown. Check the highlights before you send — they're listed for exactly that reason.
Built in
The details that make repeat practice worth it.
None of this needs configuring. It is how every interview already works.
You never see the questions
Seed questions and follow-ups stay hidden from you, during and after. You see the topics that were covered. The moment you can read the questions, a mock interview turns into a study guide.
New questions every time
Practise the same role again and Lebah writes fresh questions, having been told everything you were already asked. Same topics, same rubric — so attempt two measures skill, not memory.
Technical screens and HR rounds
Pick a technical screen, scored on depth and trade-offs, or a behavioural round, scored on whether your stories have a situation, a task, an action and a result.
English or Bahasa Indonesia
Choose the language when you create the plan. The interview is conducted in it and the report comes back in it. The interface itself switches independently.
Share a report with a link
Off by default. Turn it on and anyone with the link can read the scorecard — without your recording, your transcript or your other attempts. Turn it off and the link stops working.
Rewatch what you actually said
The interview is recorded and plays back on your report. Where the transcript is synced, clicking a line jumps the recording to that moment.
Before you upload anything
What happens to your CV, your face and your voice.
Job hunting is a bad time to be surprised by a privacy policy, so here is the short version.
Your CV file is never written to disk
PDFs and Word files are read in memory and thrown away; only the extracted text is saved, and you can see and edit that text before anything is generated from it.
The interview is recorded, and it's yours
Video and audio are recorded so you can rewatch yourself, and stored with your report. Nobody else can open it, and a report you share never includes it.
Reports are private until you say otherwise
Sharing is off unless you turn it on. A shared report carries the scorecard and nothing else — no recording, no transcript, no sign of your other attempts — and revoking the link kills it immediately.
This is practice feedback, not a verdict
Scores come from an AI reading your transcript against the posting. They are good at showing you which answers were thin and why. They are not a prediction of whether you will get the job, and no employer sees them.
Lebah is invite-only while it is small, and there is no published data-retention or self-service deletion policy yet — if you want something removed, ask and we will remove it. We publish no testimonials, user counts or success rates, because we do not have any we could honestly stand behind.
Questions
Before you start.
Everything here describes what Lebah does today.
Bring the job posting. We'll bring the interview.
One role, six steps, about fifteen minutes of talking. You will know exactly which answers need work before anyone else hears them.
Signing up needs an invite code while Lebah is small — the FAQ above explains why, and what to do if you don't have one.