Examples / HR
Founder Bar Extractor
This is not AI that screens candidates. It points at the founder: it turns “need solid engg / no one good” into a concrete, cited hiring bar — grounded in the work the team actually does, with the places the founder contradicts their own evidence surfaced. It never scores, ranks, or decides on a single applicant.
Shown on realistic sample data — an invented startup (“Northstar Robotics”) hiring a founding backend engineer. Real intake, no real candidates judged.
The leverage point
“No one good” usually means the bar was never made explicit.
A founder says hire to scale, then: “we're trying — no one good,” “need solid engg,” “all useless people mostly.” Underneath the frustration is a misalignment between what the founder actually wants and what they've written down — which is nothing. “Solid engg” is not a spec. So the job post attracts the wrong people, interviews run on vibes, and the offer surprises everyone.
The missing input that stalls all of hiring is a concrete, testable definition of the role. Make that explicit and cited, and every step behind it — the post, the screen, the interview, the offer — finally has something to align to.
So we don't build a résumé scanner. We build the thing that manufactures the missing input: the founder's real bar, pulled from their own evidence.
The build
Point it at the founder, not the candidates.
That single choice makes the build both more useful and safer.
Evidence, not opinions
The bar is pulled from what the founder can point at: the candidates they actually loved or passed on, and the work the current team really does — not a wish-list written from scratch.
Grounded in real work
Every must-have is checked against the team's actual tasks. A requirement with no basis in the work — a pedigree proxy — gets flagged, not shipped.
Their own contradictions
Where the founder wants one thing but loved a candidate who broke that rule, or the non-negotiable fights the offer, it's surfaced for them to resolve.
Never judges people
The tool structures the employer's intent and hands a rubric to a human. It does not score, rank, or decide on applicants — differentiated, and clear of the discrimination-risk of AI candidate-scoring.
How it runs
A short, inspectable chain — no black box.
Ingest with anchors
Parse the founder's intake answers, the notes on past candidates they loved or passed on, and what the current team actually does — each record tagged with its source.
Extract the bar, cited
Separate must-haves from nice-to-haves from non-negotiables, each carrying the founder's own reason and a citation back to their words. Nothing is invented.
Ground it in real work
Check every must-have against the team's actual tasks. If a requirement matches no real work, it's flagged as a possible pedigree proxy — a question, not a deletion.
Catch self-contradiction
Compare the stated bar against the evidence. A pedigree must-have against a candidate they loved without it, or an on-site rule against a remote-friendly offer, becomes a named, dual-cited question.
Stop at the human gate
Export the role bar, the candidate-facing job doc, the human screening rubric, and the open questions. The founder confirms and posts. The tool judges no one.
The test it has to pass
We write the pass/fail bar before we build.
A bar-builder that quietly invents a requirement, or drops one, or slips into ranking people, is worse than nothing. So the value isn't the demo — it's the test set. The build ships only when all six bars are green on the sample:
- Every requirement cites the founder's own words. No invented requirements.
- Coverage. Every stated intake item lands in the bar — none silently dropped.
- Must-have vs nice-to-have stay separated. No quietly-everything-is-critical.
- Ungrounded requirements are flagged. A must-have matching no real task is surfaced as a question.
- Self-contradictions surface, dual-cited. The founder-vs-evidence gaps are named, not smoothed over.
- The boundary holds. Zero candidates scored, zero ranked. The package has no such capability.
The result
Measured, not promised.
This is a real runnable pipeline, not a mockup. The figures below are its actual output over the sample intake, with a six-bar test set that has to stay green before it ships:
3 of 4 grounded
Measured: three must-haves map to real team tasks; the fourth — “5+ years at a venture-backed startup” — maps to none and is flagged as a pedigree proxy.
2 contradictions
Measured: the pedigree must-have vs a candidate the founder loved without it, and on-site-no-remote vs a remote-friendly offer — both surfaced, dual-cited.
0 candidates scored
Measured across the run: the tool builds the bar and hands a rubric to a human. It ranks nobody. Enforced by the test suite.
Every figure here is produced by the build's own passing test set on the sample intake. On a real hire we'd measure the same way — requirements grounded, contradictions caught, and the drop in wrong-fit interviews.
Why this repeats
The same skeleton fits your workflow next.
Ingest-with-citations, extract-into-a-structured-record, ground-and-check-against-evidence, prepare-cited-outputs, stop-at-the-human-gate is the same chain behind our reconciliation and quote drafter examples. Here it reconciles a founder's stated bar against their own behavior instead of money against a bank feed. Change the sources; the architecture holds. That's how a third build costs a fraction of the first — and why the example you're reading is also the starting point for yours.
Hiring, and “no one's good enough”?
Often the bar just isn't written down yet. This build starts from here — pointed at your real intake instead of the sample. Tell us the role you can't seem to fill.
Get in touch See other examples →