Hire the people who can tell when the AI is wrong.

A candidate reviews work an AI produced, with flaws planted on purpose: a pull request, a document, a design, a set of claims. You see what they caught, what they missed and what they invented, with the lines to prove it, a minute after they submit.

Start free Try a sample exercise

Free plan with five assessments a month. No card needed.

src/reservationService.tsmodifiedAssistant: looks fine to me

The sweep settles pending stock once, then expires each stale hold.

Click the line you would comment on. One of these lines is wrong.

How it works

1
Pick or generate an exercise
Start from the library, or generate one for the role in about four minutes. The generator plants the flaws and writes the answer key, and you approve it before any candidate sees it.
2
Send the link
The candidate opens it when they are ready. You decide whether there is a time limit and whether they get the assistant.
3
Read the score and decide
A draft score is ready a minute after they submit. Change anything you disagree with, pick a recommendation and finalize.

What the candidate sees

They get the whole repository, the diff, an assistant they can ask, and a verdict to give at the end. In this one the candidate has commented on a planted flaw and dismissed a claim the assistant made up.

rankrPR critiqueTime left 38:20

Make reservation transitions recoverable with per-line atomic stock apply

Ticket

src/reservationService.tsmodified

@@ -0,0 +316,16 @@
250	316
	317	+ /** Applies every still-pending line of a terminal reservation and returns the settled record. */
	318	+ private async completePendingStock(reservation: Reservation): Promise<Reservation> {
	319	+ if (reservation.pendingStockSkus.length === 0) {
	320	+ return reservation;
	321	+ }
	322	+
	323	+ const pending = new Set(reservation.pendingStockSkus);
	324	+ for (const line of reservation.lines.slice(0, -1)) {
Candidate slice(0, -1) drops the last line, so its stock change is never applied and pendingStockSkus never empties for it. A single-line reservation gets nothing applied at all. This should iterate every line.
	325	+ if (!pending.has(line.sku)) {
	326	+ continue;
	327	+ }
	328	+ await this.applyPendingLine(reservation.id, line.sku, stockChangeFor(reservation.status, line.quantity));
	329	+ }
	330	+
	331	+ const settled = await this.store.getReservation(reservation.id);

ApproveRequest changes

Request changes. The last line of every reservation is never applied, and the idempotency race path skips the payload check. The recovery test needs toBe(1). The adjustBy helper is unnecessary.

Submit review

What you see

We planted the flaws, so every exercise comes with an answer key. The score can say which flaws were caught, which were missed and what was invented.
The score is a draft until you finalize it. Every line links to the comment or prompt behind it, so you check the reasoning instead of taking a number on trust.
Candidates who took the same exercise sit side by side, so you can see where each one lands.
The assistant transcript is part of the record, so you can see what they asked it and whether they checked the answers.

Read the sample report

Final score6 / 967%

Verdict request changes, expected request changes: matched. Recommendation hire.

Plant	Outcome	Quality
p1 security (major)	Caught alone	2
p2 logic bug (major)	Caught alone	2
p3 hollow test (minor)	Caught with the assistant	1
p4 performance (major)	Missed
p5 overengineering (minor)	Caught alone	1
p6 false claim (minor)	Caught alone	2

FinalFinalized 2026-10-22 by an interviewer.

Exercises for the work the role will do

Pick the kind of work the role does day to day. The code exercises are for engineers and data people. The rest work for any role.

EngineersPR critique
Senior engineers and architectsDesign review
Data scientists and analystsVerify the claims
Product, finance, operations and peopleDocument critique
Sales and marketingRank the answers

PR critique: Review a generated pull request with planted issues, with an assistant that is sometimes wrong.; Scored: Each planted flaw: caught alone, caught with the assistant, dismissed, or missed. The verdict. Assistant claims accepted when they were false.
Challenge the AI: An AI reviewer makes confident recommendations about a codebase. Some are good, some are not. Agree or push back with reasons.; Scored: Each recommendation: the right call and the reason for it.
Verify the claims: An AI explains how the code and its libraries behave. Some claims are invented. Mark each true or false with evidence.; Scored: Each claim: true or false, with the evidence named.
Rank the answers: One question from the role's field, three AI answers of different quality. Rank them and say why. Takes about ten minutes.; Scored: The order of the three answers and the justification against the known flaws.
Design review: An AI proposes an architecture for stated requirements, as a diagram and a written proposal, with planted flaws. Comment on the design and give a verdict.; Scored: Each planted flaw in the design, and the verdict.
Rapid fire: A timed deck of short cards, each a snippet or a small diagram with one claim. True, false, or not sure, with a few seconds per card. It shows whether their confidence matches their accuracy when there is no time to check.; Scored: Each card: right, wrong, or not sure. Confident and wrong costs more than not sure.
Document critique: An AI wrote a document from a request and source material: a plan, an analysis, a proposal. Some of it is wrong. Comment on the document and give a verdict, with an assistant that is sometimes wrong.; Scored: Each planted flaw in the document, the verdict, and the assistant claims accepted when they were false.

Fits the tools you already run

API

Bearer keys issued in Settings. List items, add candidates, create sessions and read scores from your own tooling.

GET /api/v1/items
POST /api/v1/candidates
POST /api/v1/sessions
GET /api/v1/sessions/{id}

Webhooks

Signed with a per-endpoint secret and retried with backoff, so your pipeline hears the moment something changes.

session.submitted
session.scored
score.finalized

Applicant tracking

People have asked for Greenhouse, Lever and Ashby connectors, and they are not built yet. The API and webhooks cover the same steps today. Tell us which one you need and we will build it with you.

Prices on the page. No sales call.

Pay by the month for assessments, generated items and seats. It costs less than every other assessment platform we know of, and a yearly term gets two months free.

See the full comparison

	Free	Starter	Team	Business
A month	$0	$29	$69	$199
Assessments a month	5	25	100	400
Generated items a month	2	10	40	150
Members	3	5	20	Unlimited
Verified email domains	No	No	Included	Included
	Start free	Start with Starter	Start with Team	Talk to us

What happens to candidate data

Kept inside your organization

Every query is scoped to your organization. Candidate text is fenced before any model sees it, and no provider sees a CV during scoring.

Export and erase

Export everything held about a candidate as a file. Erase them in one click, with the CV, sessions and scores.

Retention you set

Choose how long closed candidates are kept. After that they are erased automatically and the erasure is logged.

A signed DPA in the terms

Processor terms are part of the terms of service, so there is nothing to negotiate before the first candidate.

SOC 2 is on the roadmap

We will publish the audit window once it is scheduled. Until then the security page says exactly what is in place, and what is not.

Why we built it

Rankr started as the exercise Gearlay gave its own candidates. Once everyone arrived with an assistant, take-home tests and portfolio code stopped telling us anything; what separated people was whether they noticed when the assistant was wrong, and whether they could say why. We built that exercise, then the scoring, then the rest of the product around it. There are no customer logos on this page yet. We would rather be judged on the exercise.

Run your first assessment this afternoon.

Start free

Business plan or more than one organization? Talk to us.