EvidenceHire · The product
A deep-AI system for finding out what people actually know.
EvidenceHire turns unstructured material and live conversation into a decision record a person can inspect. Recruiters use it to verify role fit. Educators use it to rehearse a whole class for placement — on the material they actually taught.
Module 01 · Hiring intelligence
Screen on cited evidence, not self-reported claims.
EvidenceHire reconstructs a candidate's professional footprint from every source they've left behind, cross-checks each claim against it, and returns a ranked fit report where every verdict quotes the excerpt it came from.
A profile rebuilt from every source, not the one they sent you
LinkedIn, résumé PDFs and DOCX, GitHub, Dribbble, Behance, Upwork, personal portfolios and publications. Any link found inside a résumé is fetched too. Everything is content-hashed, so the same evidence is never paid for twice.
Per-skill verdicts with the excerpt attached
Every claimed skill is checked on its own and comes back corroborated, weakly corroborated, claim-only or contradicted — with depth, recency, a confidence score, and the verbatim excerpt and document each verdict rests on.
Skills checks generated for that candidate's gaps
20 to 30 open scenario questions built from the role's must-haves, the candidate's weak claims and the topics you want probed. Free text only. Each answer is graded against a per-question rubric and screened for AI authorship.
A live AI interview when the evidence is thin
A one-time link with no sign-up opens a spoken conversation with an AI interviewer that follows up on the candidate's own background. Hard time cap, one attempt, recording and transcript kept, plus an AI-authorship signal and a vision-based real-person check — all advisory.
Flight risk and recruiter intel, before you make the call
Retention likelihood with its drivers and counterforces. Then outreach templates, an opening hook, questions to ask, what to listen for, compensation estimates and negotiation style.
Built for the volume agencies actually run
Bulk CSV import of a LinkedIn Recruiter export, public apply links that are rate-limited and blacklist-aware, bulk assessment assignment above a score threshold, and branded shortlist PDFs. Re-screening the same person against a new role reuses their profile, so it costs a fraction of a fresh analysis.
How the recruiter agents work
A chain of narrow jobs, with evidence handed from one to the next.
The product does not ask one model to decide whether someone is hireable. It separates collection, normalisation, claim checking and role-fit work, so each conclusion has a defined input and a recruiter can see where uncertainty entered.
- 01
Evidence collector
Pulls the résumé, LinkedIn and the public links the candidate supplied or embedded in their documents. Each source is extracted, labelled and content-hashed before it is used.
Output · A reusable evidence pack, with the original source identity preserved.
- 02
Profile normaliser
Turns the raw material into structured roles, dates, education, publications and claimed skills. It is instructed to attach the exact source excerpt instead of silently filling gaps.
Output · A structured profile whose fields can be reopened to their source.
- 03
Claim-verification specialists
Run a separate evidence check for each claimed skill, then read leadership, mobility and tenure patterns as their own bounded questions. A weak or missing claim stays weak or missing; it is not upgraded because the résumé sounds polished.
Output · Per-skill status, depth, recency, confidence and 1–5 supporting evidence pointers.
- 04
Role-fit and challenge designer
Compares the verified profile with this role's must-haves and nice-to-haves. Where evidence is thin, it writes open, candidate-specific scenarios or drives a spoken interview with adaptive follow-ups.
Output · A ranked, role-specific report and fresh evidence from answers the candidate actually gave.
What the recruiter data looks like
A score is the summary. The evidence is the record.
The shortlist is designed to be read in either direction: scan the ranked score first, or open a single skill and trace it to the source, date and assessment result that informed it.
- Source traceability
- Every source-derived field and verdict retains a verbatim excerpt or evidence pointer; a recruiter can inspect why it exists.
- Explicit uncertainty
- A claim is labelled corroborated, weakly corroborated, claim-only or contradicted. Depth, recency and confidence sit beside the label instead of being hidden inside one score.
- Role-specific comparability
- The same deterministic weighting calculates every role score: skill match 40%, depth 20%, leadership 15% when relevant, relocation 10% and retention 15%. If leadership does not apply, its weight is redistributed rather than treated as a zero.
- Controlled reuse
- Content hashes and input fingerprints reuse unchanged evidence for a new role, while a changed document, assessment result or prompt version triggers a fresh relevant analysis.
The verification pipeline
Ten specialised stages, not one generic prompt.
Each stage does one job and cites its evidence. Results are cached per candidate, so re-screening the same person against a new role reuses everything that has not changed.
- 00
Ingest
LinkedIn profiles, résumé PDFs and DOCX, and every external link found inside them are fetched and content-hashed for reuse.
- 01
Normalise
Raw evidence becomes a structured profile — roles, dates, education, publications, claimed skills. Every field cites the exact excerpt it came from; nothing is silently inferred.
- 02
Skill evidence check
Each claimed skill is checked individually against all available evidence and classified corroborated, weakly corroborated, claim-only or contradicted, with verbatim evidence pointers, depth and recency.
- 03
Leadership signal
Team size managed, span of control, budget and P&L ownership, with narrative examples and citations.
- 04
Relocation fit
Whether the candidate's location actually works for this role — metro, country, remote compatibility, and how willing they signal they are.
- 05
Flight risk
Retention likelihood from tenure patterns, career trajectory and market demand, with drivers and counterforces stated.
- 06
Fit assessment
Skill-by-skill matching against the job description's must-haves, combined into a weighted score: skill match 40%, depth 20%, leadership 15%, relocation 10%, retention 15%.
- 07
Recruiter intel
Outreach templates, interview questions to ask, what to listen for, compensation estimates and negotiation-style analysis.
- 08
Skills check
Where evidence is thin, 20–30 open scenario questions are generated for that candidate's specific gaps. Never multiple choice — there is nothing to guess between.
- 09
Live AI interview
A spoken conversation with an AI interviewer that probes the doubtful claims, with recording, transcript, AI-authorship signal and a vision-based real-person check.
Module 02 · Placement readiness
Put a whole class through the interview before the panel does.
The same live interviewer that verifies a candidate's claims for a recruiter runs mock interviews for students. Pointed at your syllabus instead of a job description, it puts an entire class through a timed spoken interview built from the material you taught.
01
Describe the course, attach what you taught
Course name and code, department, the year or semester as your institute writes it, and a brief curriculum. An ebook, lecture notes or a case pack can be attached; its text is read alongside the curriculum, not instead of it, and links inside it are fetched as extra grounding.
02
Topics drafted, then edited by faculty
Six to twelve interview topics are drafted from that material — what panels in the discipline actually probe, which is deliberately not the book's table of contents. A faculty member reviews and edits them before anything is published.
03
Publish seats, difficulty and a window
You set the number of seats, the difficulty and when the session is open. A seat is claimed only when a student actually starts, and unused seats are refunded when the session closes.
04
One class link, one attempt each
Students enter a registration number and a name, then start. No account, no install. The one-attempt rule is a unique index in the database, not app logic that can be talked around.
05
Graded on substance, reasoning, and what to fix
Per-topic scores with a note on each, a separate problem-solving read, and at most three strengths and three improvements. Faculty can override a grade outright, and prior versions are kept — a changed grade with no trail is the one that gets disputed.
06
Reports at every level you need them
A one-page report per student with radar, gauge and transcript. A ranked class score sheet as PDF or CSV. Cohort insights per department and term, an at-risk list, per-faculty cohorts for the head of department, and one student across semesters with a trend line.
How the educator agents work
A faculty-controlled session, followed by evidence a teacher can use.
The educator module uses the same live interview engine, but its job is different: turn the material you taught into placement practice, protect a fair session, and give faculty a specific record of each student’s readiness.
- 01
Course-to-panel planner
Reads the course details, curriculum and optional ebook, notes or case pack, plus relevant links found in them. It drafts 6–12 applied topics that a viva or placement panel would probe — not a copied table of contents.
Output · A faculty-editable session brief, discipline read and ordered interview topics.
- 02
Live interviewer
Runs one short, spoken conversation per student. It asks one question at a time, follows the student's actual answer into a concrete how, why or trade-off, and keeps acknowledgement neutral so it never reveals whether an answer landed.
Output · A time-bounded recording and transcript of what the student demonstrated.
- 03
Transcript evaluator
Grades each published topic from the student's own spoken answers, then separately reads problem-solving: how they broke down an unfamiliar question, applied a principle or corrected their thinking. It rewards substance, not polish, accent or length.
Output · A per-topic record, knowledge score, problem-solving score, short strengths and concrete improvements.
- 04
Faculty reporting layer
Builds the student report, class sheet and cohort views from the same attempt records. It keeps the transcript and recording available for review, and lets faculty override a score with a reason without overwriting the original grade.
Output · Traceable student, class, department and term-level readiness data.
What the educator data looks like
A cohort view that never loses the student-level evidence beneath it.
Every number in a class, department or term view comes from the same per-student attempt. Faculty can review the transcript and recording, then make an accountable correction when their judgement differs.
- Grounded before it is generated
- Faculty start with what they taught and approve or edit the topics before publishing. The agent may frame the material around current panel practice, but it cannot silently replace the course.
- One comparable attempt
- A student is identified by registration number, and the one-attempt rule is enforced in the database. Session timing, transcript and submitted record make cohort rows comparable.
- Substance separated from delivery
- Knowledge comes from the interview answers; problem solving is measured separately. Optional voice, expression and body-language coaching never changes the marks earned for subject knowledge. An unmeasured axis is null, never mistaken for zero.
- Reviewable, not automatic discipline
- Any integrity read is advisory and can be hidden for an institute. It is not a finding of misconduct or an automatic decision; faculty can review the underlying recording and retain an override trail.
For the student
Job-ready is a thing you can practise, and then measure.
Students are never shown a score — the interviewer is built never to signal how an answer landed. The report goes to the faculty member, who decides what it means and what to say.
The real interview, not a lighter version
Students sit the identical spoken AI interview a candidate sits. It follows up on what it hears, so a rehearsed answer comes apart exactly the way it would in front of a panel — in a room where that costs them nothing.
Answers graded on substance and on reasoning
Per-topic scores say what the student knew. A separate problem-solving read says how they reasoned, which is the part a viva or a case round is actually testing.
At most three strengths and three improvements
Specific enough to act on before the next round. Not a wall of generated feedback nobody reads.
Five competency axes and a placement-readiness number
Knowledge, problem solving, communication, confidence and body language, blended into one readiness figure computed the same way every time — so a trend line across semesters actually means something.
Delivery coached, never penalised
Where an institute enables it, the recording is read a second time for verbal, voice, expression, body language and behavioural feedback. It never moves the score: a nervous student who knows the material keeps every mark their answers earned.
Graded on the material, not on the accent
The grader carries explicit fairness rules — never score down for accent, dialect, a disability, a speech difference, appearance, or a poor camera and room.
Five axes, one placement-readiness number
Computed the same way every time rather than asked of a model, so it stays comparable across sessions, semesters and product versions. An axis that was not measured is null, never zero — a zero would read as “did terribly.”
- Knowledge
- The interview score — what the student knew about the material.
- Problem solving
- How they reasoned, as opposed to what they recalled.
- Communication
- Verbal clarity, weighted with voice.
- Confidence
- Voice, expression and behavioural signals blended.
- Body language
- Read from the recording, where the institute enables it.
For the institution
What a university, a department and a placement cell each get.
The same interview your students will face, run as a rehearsal you control — months before the placement season starts.
Placement readiness you can measure before the season
Run a class in week one and you know which students are ready, which are close, and which need work — months before recruiters arrive on campus.
Topics from your syllabus, not a fixed question list
The session reads your curriculum and course material, then drafts what panels in that discipline are actually probing — deliberately not the book's table of contents. Faculty review and edit before anything is published.
Every discipline, not just engineering
Law, management, architecture, journalism, engineering — the topics come from the material, so the module does not care which department it is pointed at.
It runs on your department structure
Sessions carry the course, the code, the department and the year or semester as your institute writes it. The head of department sees every session, faculty see their own, and results group the way your records already do.
Cohort numbers that trace back to a name
Which topics the class is weak on, who scored under fifty, how one student moves across terms. Every cohort figure is built from the same student reports, so it can always be opened up and traced to a person.
Practice, not proctoring
No lockdown browser, no recording of a student's own machine, no facial recognition and no biometric template. Nothing here works out who a student is. Faculty can override a grade, and the trail of what changed is kept.
Where we hold the line
Both modules judge people. That sets the rules.
No protected attributes
Gender, race, age, religion, disability and profile photos are excluded from scoring and ranking. Only job-relevant evidence counts.
The human makes the call
Every report is decision support. The real-person check on a live interview is advisory and never an auto-reject.
Tenant-isolated, GDPR-aligned
Every query is scoped to your organisation, and documents and recordings live in access-controlled storage.
Auditable by design
Verbatim citations, per-stage run logs and a faculty-override trail. Anything the system concluded can be reopened.
Want a closer look?
Run it on one of your roles, or one of your classes.
Also from ChitraYantra
VidiMitraLegal research
Indian legal research, grounded in authority. Our second product: grounded research and drafting for Indian law, with page-pinned authority, balanced research memos and lawyer review at every consequential step.
- Grounded case & precedent research
- Page-pinned authority trails
- Lawyer-reviewable drafts