Rubric you approve
Nothing is scored against criteria you have not read. Regenerate with an instruction; approve to lock a version.
AI hiring pipeline
Every CV scored. Every score explained.
HRAgent turns a job description into a frozen rubric, scores every applicant with a 3-vote model ensemble, verifies claims against public evidence, and hands you the same top 10 every time you run it.
No demo data in the product. What you see is what ran.
5/5identical top-10 sets across isolated runs
3xindependent scores per criterion
100%of model calls stored verbatim
0hardcoded evaluation rules
How it works
From job description to defensible shortlist.
Approve the rubric
Your JD becomes weighted criteria. The run freezes the version you approve.
Collect and screen
A hosted form validates, virus-scans, and de-duplicates every CV.
Score three times
Independent votes per criterion, median wins. Public claims go out for verification.
Cut the top 10
Deterministic aggregation, pairwise checks at the cutoff. Rank 11 is explainable.
How it works
From job description to defensible shortlist.
Approve the rubric
Your JD becomes weighted criteria. The run freezes the version you approve.
Collect and screen
A hosted form validates, virus-scans, and de-duplicates every CV.
Score three times
Independent votes per criterion, median wins. Public claims go out for verification.
Cut the top 10
Deterministic aggregation, pairwise checks at the cutoff. Rank 11 is explainable.
What you get
Evidence on every score
Each 0 to 4 score carries verbatim quotes from the CV. No score exists without its receipt.
Background verification
Public profiles are checked out of band. Verified, unverified, or contradicted, with sources.
Live run monitor
Watch every candidate move through parsing, scoring, and verification as it happens. Real events only.
Immutable audit ledger
Every model call, vote, and comparison is stored. Overrides are recorded, never overwritten.
Isolated by design
Row-level security per tenant, signed webhooks, virus-scanned uploads, erasure on request.
Live monitor
Watch the pipeline think.
This is the actual run monitor component running a scripted 12-candidate scenario.
Simulated preview
Queued12
Parsing0
Scoring0
BG check0
Verifying0
Done0
Consistency
Same CVs. Five runs. One shortlist.
Ensemble medians, integer-only aggregation, and pairwise checks at the cutoff. We gate releases on five isolated runs producing an identical top-10 set, with no caching.
Run 1
C-04A7C-1B32C-9E01C-77C4C-D218C-30AFC-6F55C-A9E3C-52B8C-E76D
Run 2
C-04A7C-1B32C-9E01C-77C4C-D218C-30AFC-6F55C-A9E3C-52B8C-E76D
Run 3
C-04A7C-1B32C-9E01C-77C4C-D218C-30AFC-6F55C-A9E3C-52B8C-E76D
Run 4
C-04A7C-1B32C-9E01C-77C4C-D218C-30AFC-6F55C-A9E3C-52B8C-E76D
Run 5
C-04A7C-1B32C-9E01C-77C4C-D218C-30AFC-6F55C-A9E3C-52B8C-E76D
Jaccard(top-10) = 1.0
Verification
Claims are checked, not trusted.
Candidates with public links get an out-of-band background check. Findings come back as condensed, sourced summaries; the verifier can adjust any score by at most one point, and both numbers are kept.
4 years maintaining a public Python ETL library.
github.com/…/etl-kit
unverifiedverified (github)
Security
Boring where it counts.
Tenant isolation
Postgres row-level security enforced in the database.
Screened uploads
Magic-byte checks and ClamAV scanning, fail closed.
Signed webhooks
HMAC in both directions with replay protection.
Right to erasure
Candidate data deleted on request; the audit structure survives without it.
Run your next hire through it.
Onboarding is guided: bring a JD, approve a rubric, share a link.