One hiring conversation, four sides of the table.

Screening at a volume no team could staff. It runs the first-round conversation on every applicant instead of the twenty someone had time for, and hands back a ranked shortlist rather than a call queue.

This is the end of the platform that runs on the phone, on real numbers, inbound and outbound — so the constraint is not the model, it is the line. It is already talking as the candidate says hello, because a silent beat on answer is the pause that makes people hang up on an unknown number. Answering-machine detection ends the call rather than interviewing a recording.

The ranking carries its reasoning. A shortlist a recruiter cannot interrogate is a shortlist they will re-do by hand, so every verdict arrives with the answers it was drawn from and the requirement it was judged against — and it reaches the recruiter’s systems as structured fields, not a transcript to read.

The candidate points the system at the actual job they want and gets a straight answer: where they already match, where they are weak, and what closes the gap. Not improve your communication skills — either this fact is in your resume but buried on page two, which is an edit, or this requirement has no evidence anywhere, which is a project.

Two scores, never blended. Fit asks how much of what the job wants the resume’s substance actually satisfies. Conveyance asks how findable that substance is. Two resumes with identical facts, one crisp and one badly structured, have the same fit and different conveyance. A single blended number would hide exactly the thing the candidate needs to know.

A resume is a repository, not a file. Every change is an immutable version; it can branch into a data-analyst line and a devops line that evolve independently; any two versions diff line by line, and one is pinned as the one that actually goes out. Gaps marked missing generate a scoped project with milestones, tracked by hand with an artifact attached, that lands back in the resume as a reviewable diff.

A final-year student in a placement programme alongside three hundred classmates. The traditional version is a set of workshops, a couple of mock interviews if they are lucky, and no idea how they would actually perform under pressure.

It is the same practice engine as the medical and frontline training — the avatar that plays the anxious patient for Singapore’s universities and the resident who has decided no for its ministries — pointed at a job interview instead of a consultation. Not a similar idea reimplemented: the same engine, already in production in rooms with more at stake. See it running

Here that becomes a live, AI-run interview that behaves like the real thing: questions that adapt to the answers, follow-ups that probe when something sounds thin, and feedback afterwards naming exactly where they stumbled, turn by turn, with the video to prove it. The student picks which resume and which versionthey are presenting before it starts, and that version is pinned to the session permanently — so the score beside the session and the words spoken in it refer to the same document. Your paper says X and you said Y about X out loud are finally comparable.

The room shows the agent’s own state — waiting, listening, thinking, speaking — because a silent agent that is thinking and a silent agent that has crashed look identical otherwise. The whole session records: webcam video plus a two-channel mix, agent on one side and student on the other.

Not another dashboard. What an educator gets is a short list of names: the students where coaching demonstrably failed to land, so a human can step in. Students with nothing stuck are omitted entirely — it is a to-do list, not a roster. That is also the answer to does this replace teachers: no, it hands them the people it could not reach.

Around that sit the numbers a placement cell actually decides on. A requirement × class heat map, where 212 of 300 fail production deployment experience is a curriculum decision. Class trajectory plotted by session ordinal rather than by date, because students start on different days and a calendar series measures scheduling, not learning. Bail-outs counted separately, because a three-turn session is not a short session, it is someone who quit, and counting it as completed flatters every other number on the page. And drive readiness: before a company visits, how many eligible students clear its bar.

The teacher decides what the interview is for. Before a session runs, they set what it should lean on for a particular student or a whole class — the requirement a cohort keeps failing, the thing this one person needs to be pushed on. It is not a difficulty slider: it changes what the agent asks about and what it lets go.

That is also where accommodations live, and they are behavioural rather than cosmetic. An autistic student can be marked so the agent does not probe eye contact and does not counter them on it — the agent stops doing the thing, rather than doing it and then discounting the result. And the student picks push-to-talk if they want it, so a stammer or a pause for thought never ends their turn early and costs them the interview. Nobody should lose a round to the microphone.

The machine never has the final word about a person. Readiness is computed as an advisory band and then has to be conferred by a named human who saw the basis for it. Two fields, never one; the default is not-ready, and a computed band moving to ready promotes nobody — it moves them into a review queue.