Recruit reads a resume, cites every claim back to its source, cross-checks the evidence, and interviews what's still unproven — then hands a human a decision they can defend months later.
No scores. No black-box ranking. Every verdict cited; every decision human.
github.com/…/k8s-operator — 3 yrs of commits, cited
Corroborated by title; scope unconfirmed — flagged for interview
Overlaps a full-time MS in the same window
Four stages turn a PDF into a defensible hiring decision. Every stage is offline-testable and leaves an auditable trace.
Extract every claim
The resume is parsed into discrete, verifiable claims — each one cited back to the exact sentence it came from. No paraphrasing, no invented detail.
Cross-check the evidence
Claims are checked against public signals — GitHub, package registries, papers, patents — while consistency checks catch impossible timelines and overlaps.
Interview what's still unproven
For claims evidence can't settle, the AI generates adaptive questions grounded in the candidate's own experience. Vague answers get probed, not passed.
Deliver a cited verdict
A hiring summary where every verdict links to its evidence. No score, no ranking, no automated reject — a human makes the call, fully informed.
The hard guarantees live in code, not in a prompt — which is exactly what makes the output defensible.
Evidence it can't fake
The citation guardrail is enforced in code: a source is only accepted if the URL resolves and the quote is a literal substring of it. The model physically cannot cite something that isn't there.
A tamper-evident record
Every step — extraction, evidence, interview, decision — lands in a hash-chained Evidence Ledger. Months later, you can prove exactly how a call was reached.
Inbound that verifies itself
Share one public apply link. Candidates apply with their own resume and verification starts the instant they submit — zero data entry, every applicant pre-checked.
Human-in-the-loop by design
No scores. No rankings. No automated pass/fail. The system surfaces cited evidence; a person decides — auditable, defensible, and bias-conscious.
Scored against a hand-labelled dataset of real resumes, planted lies, and messy edge cases. The number that matters most: how often a fabrication slips through as verified.
98.4%
Claim-extraction F1
resumes → clean, citable claims
100%
Citation validity
every cited span is real
0
Planted lies passed as verified
on the fabrication-safety set
Paste a job description, drop in resumes or share an apply link, and let the evidence pipeline do the first pass — so your team spends its time on the candidates who hold up.