How it works
From a job posting
to a first-round decision,
in one place.
Create a job, take applications, read what arrives. The candidates you choose take an AI interview, and their answers come back as a report.
Everything up to deciding who is worth meeting.
- Jobs
- Applicants
- AI interview
- Reports
All four live in one console.
The whole loop
The first round, joined up.
This is how an application becomes a decision. Each stage below is taken in order, beside the screen it happens on.
- 01JobCriteria and questions written
- 02ApplicationsThrough the job's public link
- 03DocumentsWhat is shown, and what is not
- 04AI interviewFor the candidates you chose
- 05ReportAnswers and the evidence, together
- 06A person decidesAnd meets the people worth it
What Hitoshiru puts in order
What a person decides
Hiring is not handed to the AI.
It puts the evidence in order; a person decides.
Whom to meet and whom to hire are decided the way they always were. What Hitoshiru does is bring the material for that decision to the same standard for every candidate.
The job, and its questions
Write the job, and the questions arrive.
The model writes question suggestions from the posting and the criteria. None of them is used as it stands: a recruiter picks, edits, and publishes, and only then does a question reach an interview.
- 01
Create the job
The role, the work, what the interview is for, and the criteria with their weights, on one screen.
- 02
The model suggests questions
Written from the posting, the criteria and what the interview is for. The suggestions are stored nowhere, and any one question can be replaced on its own.
- 03
Review and edit them
Only what you tick is added to the job. It is used in an interview once you publish, and not before.
4 suggested questions
AI suggestions, written from this job's description.
Nothing is added until you tick it. These are written from the job description and your scoring criteria, and each one was screened against the 職業安定法 14 topics.
Tell us about winning a new account — from the first contact through to the contract, as concretely as you can.
For New business
If there was a period when you missed your number, what did you conclude the cause was, and what did you change the following quarter?
For Meeting a target
When a customer wanted something that cut against your own company's policy, how did you settle it?
For Coordinating inside and out
When you put a proposal together, what do you find out about the customer's industry, and how far do you go?
For Understanding the customer
1 more was withheld by the 職業安定法 screen.
The console's suggestion panel, drawn. This is a moment mid-review; nothing is ticked when the panel opens.
Applications
Publish the job, and take the applications here.
Every job gets its own application URL. Put it on your careers site, in a post, or in a job advert, and candidates apply straight from the job page. The form takes one CV alongside the answers.
- 04
Publish the application link
One link per job, for your own careers site, social, or a job board.
- 05
Candidates apply
Their experience and one CV arrive through the form. Intake closes on its own when the job's cap is reached.
Application link
A link for your own careers site, a message to a list, or a printed advert. Anyone who opens it can apply to this job.
https://acme.hitoshiru.com/en/apply/7f3a…This job's application link is live.
Posted on: careers page, social, job boards
Name
Misaki Sato
Email address
m.sato@example.com
About you / your experience
Four years selling to companies, mostly working with…
CV / résumé (optional)
misaki-sato-cv.pdf
I have read the above and agree to my information being used for this application.
B2B sales · 3 applications
- Move to interview
Misaki Sato
Four years selling to companies · CV attached
- Hold
Ken Takahashi
Moving from consumer sales to selling to companies
- Not handled
Nguyen Thi Lan
Sales support at a trading company, Japanese N2
Only the people you move to interview are sent an interview link.
The job editor, the application form and the applicant list, drawn. Not a real company or real applicants.
Documents, and who goes forward
Who gets an interview is yours to decide.
Read what arrived
The model reads the attached document against the job's criteria and sets out, per criterion, what it shows, what it does not, and what to confirm in the interview.
There is no score. A document is the candidate's own account of themselves and does not carry the weight of something they said in an interview. Every sentence read out of it is screened against the Employment Security Act's 14 topics, and anything that matches is not shown.
Only the people you move to interview are sent one. Hold and reject are recorded too, with the reason.
Misaki Sato
misaki-sato-cv.pdf
The model's reading of the document
Four years, mostly winning new business from companies. The industries and the number of accounts are stated; the last two years read as work on existing customers.
- New businessshown
Four years winning new business from companies, with the number of contracts a year.
- Meeting a targetpartly
States the attainment against target for the last year.
Not shown: Nothing about what changed after a quarter that missed.
- Coordinating inside and outnot shown
Not shown: No account of working with other departments.
- Understanding the customerpartly
States the industries covered and the sectors of the main accounts.
Not shown: Nothing about how the groundwork before a proposal is done.
To confirm in the interview
- In a quarter where you missed the target, what did you conclude the cause was, and what did you change the quarter after?
- If you have run something that needed another department's cooperation, tell us how that went.
A model's reading of the document. Not a score and not a verdict — whether to interview is yours to decide.
The console's applicant screen, drawn. Not a real applicant.
The AI interview, from the candidate's side
An interview that is easy to take.
Send it, and the candidate takes it
Everyone you move forward gets their own link by email. Nothing to schedule.
No app to install and no account to create. They open the link in a phone browser, check the microphone and the connection, answer a practice question, and begin. The camera is never used; what is on screen is our AI interviewer.
- No app
- Works on a phone
- Any hour, until it expires
- A thin answer is followed up on the spot
- An answer that ends at "I worked hard at it" gets a follow-up asking for the actual situation. That follow-up is screened against the 14 topics before it is spoken, like every other question.
- A pause for thought is not interrupted
- Hesitating and finishing are told apart. Someone who stops mid-thought is waited for; the interview moves on only when the words do not come.
- Skipping a question means being asked again
- Whether the question was not heard or not wanted cannot be told apart in the moment. It is rephrased and asked once more; if there is still no answer the interview moves on, and the report says so.
- The interviewer's voice never changes mid-interview
- One interview is one voice, start to finish. A voice that changes half way through reads to the candidate as being handed to somebody else.
- A dropped connection resumes where it stopped
- The interview picks up from the last answered question. Nobody has to start again.
The candidate report
You do not have to watch the interview.
This is what a recruiter reads.
Read the report
Every criterion's score carries the words behind it: the original, and a translation into whatever language the job writes its reports in. Any score can be overwritten with a reason. The sample below is not a real candidate.
Three years at a special nursing home, with hands-on care and night shifts, and examples given concretely throughout. What stands out is being able to describe a mistake and the correction that followed.
1Strengths
- Hands-on care and night-shift experience
- Describes a mistake and the fix in their own words
2Areas to improve (what to confirm when you meet)
- Few concrete examples of records and handovers
- Only one example of working across professions
3Bar length is the score, thickness is the weight. The line is the median for this role.
- Communicating with residentsStrong4.5
Weight 22% · role median 2
- Care experienceStrong4.5
Weight 33% · role median 1.5
- Safety and hygieneStrong4.0
Weight 22% · role median 1.5
- Working as a teamStrong4.0
Weight 22% · role median 1
Japanese proficiency (estimated from the conversation)
JLPT does not test speaking, so this is shown as an equivalent band.
CEFR (speaking)
JLPT equivalent
N2–N1 equivalent
Confidence
Medium
Assessed from 1 min 55 s of speech. Confidence is Medium because there was not much of it.
4Quoted evidence — 12:04
「最初は止めようとしましたが、余計に興奮させてしまって。それで先輩に相談したら、一緒に歩いてみたらどうかと言われました。」
“At first I tried to stop her, but that only agitated her more. So I asked a senior colleague, who suggested walking with her instead.”
A recruiter can overwrite any score with a reason. The rubric version, the prompt-pack hash and the model version are written to that interview's audit log.
The strengths
What the candidate actually evidenced, restated in a line each.
What was thin, and what to confirm
Where the interview did not reach. It is also the list of what to ask when you meet.
The rubric and its weights
Each criterion the job set, scored 0–5. Bar thickness is the weight.
The evidence for a score
The candidate's own words behind the number, quoted rather than summarised.
The full transcript and the audio are on the same screen. How Japanese proficiency is assessed is on its own page.
What the recruiter does
Four things, and only four
- 01
Review the job
Check and edit the suggestions
- 02
Choose who to interview
Read the application and the CV
- 03
Send the AI interview
One click per candidate
- 04
Read the report
Decide who is worth meeting
Hitoshiru carries the repetition around them.
How we think about it
We have decided what not to leave to the AI.
Hitoshiru is there to support a recruiter's judgement, not to reach a hiring decision on its own. The three below are not settings that can be enabled; they are decisions, and each one carries the basis it rests on.
Facial and emotion analysis
A facial image is a personal identifier code, and an evaluation report does not need one. Video is off by default and audio is all that is handled.
Basis: APPI
Questions touching the 14 topics
Every question is screened mechanically before it is spoken. A question that matches is discarded and never reaches the candidate — and the discarding is itself recorded.
Basis: Employment Security Act
Editing the record afterwards
An append-only log, hash-chained. Rewrite one line and every line after it fails verification.
Basis: Specification (append-only)
The question screen is set out on Employment Security Act, and how recordings are handled on Security.
The record
Keep the whole of how it was decided.
- The questions
- Who was asked what. Including the ones that were discarded.
- The answers
- How they answered, as transcript and as audio.
- The scores
- Why each score is what it is — and the reason for any override.
Nothing here can be rewritten afterwards.
The consent version, the rubric version, the model version, the scores and their justifications, and even the questions discarded before they were spoken — one row each. Each row's hash is computed over the row before it, so rewriting one line in the middle makes every line after it fail verification.
A sample of the log below, not a real interview.
Audit log
Append-only, hash-chained
#41consent.grantedConsent was recorded
Consent text v3, with the time and the version
hash 9f2c…a41d#42session.startedThe interview started
Rubric v7, prompt pack 8b31…
hash c084…7e12#43probe.blockedA follow-up question was discarded
Matched the Employment Security Act's 14 topics (family)
hash 1ab7…03f9#44evaluation.scoredThe evaluation was written
Model version, scores and justifications
hash 77de…b5c0#45evaluation.overriddenA recruiter overwrote a score
A reason is required; the original value is kept
hash e5a1…9d44
Each row's hash is computed over the row before it. Rewrite a single line and every line after it fails verification.A deletion does not break the chain.Only the content is erased; the hash stays, by design.
Security
Recordings are stored encrypted.
- Recordings stay on our own servers
- No third-party object storage is used.
- A separate key per recording
- Keys are held apart from the audio, and backups carry ciphertext only.
- Interview links are never stored
- Only a hash is kept. A link that has been sent cannot be read back, even by us.
Speech recognition, question generation, speech synthesis and the report are done through OpenAI's API, and that processing happens outside Japan. The consent screen before an interview names the processor. We do not print certifications we do not hold. The detail is on Security.
See the real thing
Bring a job posting and we will draft its criteria on the call. About thirty minutes covers creating the job, the candidate's screen, and the report a recruiter reads.