Grade one submission by hand. Then flip the switch
Below is a coding assessment from a hiring round. Score it the way you do today. Then hand the same console to the agent and watch it grade the identical file, run the hidden tests, find the off-by-one on a specific line, and score every skill against the same rubric. Evidence, not gut feel.
Free credits on signup. No demo gate. Metered per assessment.
Score one file by hand, then let the agent grade the identical file against the hidden tests.
Mark twelve items with a ruler, or read the auto-marked key against the same answers.
A messy tracker sheet, or a ranked evidence scorecard with per-skill bars on the same rubric.
Sample data. Every name, answer and score on this page is invented for the story. See it on your own roles.
Scroll to explore
Act 1 · Station 01 · the code submission
Same file, same rubric. Only the verdict swaps
Candidate note. Kept it in memory for simplicity. Eviction happens lazily on read.
Submitted in 18:40 of a 45:00 window · 2 test runs before submitting.
- Sliding window evicts expired hits20
- Boundary behaviour at exactly the limit25
- Per-key isolation under mixed traffic15
- Memory with many short-lived keys15
- Naming, structure, readability10
- Edge cases considered and tested15
Keep the rubric weights consistent across all thirty-eight. It rarely holds by paper thirty.
- ✓test_allows_under_limit99 calls in a 60 s window, all allowed.
- ✕test_boundary_exact_limitExpected 101st call rejected. Got 101 allowed. Line 17.
- ✓test_window_expiryHits older than the window are evicted on read.
- ✓test_keys_isolatedTwo keys at 90 calls each stay independent.
if len(q) > self.limit should be >=. The window admits limit + 1 calls. Everything else in the class is correct. This is the one character the tests care about.
Tab left the assessment 0 times. One paste of 12 characters matches the import line. Two test runs from inside the editor.
Signals, not a verdict. Nobody is disqualified by a machine. The flags go to you.
Read twenty lines, tick six boxes, write a note. Then do it thirty-seven more times.
The same file, graded on submit. Three of four tests pass, the off-by-one is on line 17.
Act 2 · Station 02 · the MCQ sheet
The part you mark with a ruler
Section 1 of the same paper: twelve items, six shown. The candidate answers are identical. Only the mark column swaps.
| Item | Answer | Mark |
|---|---|---|
| Q1. Which index lets this query range-scan on created_at? | C | ✓✕✓ correctKey C. 1.0 / 1.0 |
| Q2. A read replica lags by 4 s. Which read must hit the primary? | B | ✓✕✓ correctKey B. 1.0 / 1.0 |
| Q3. Which HTTP status should a rate-limited client receive? | A | ✓✕✕ wrongKey D (429). Picked 503. 0 / 1.0 |
| Q4. Two workers pop the same job. Safe fix without a lock table? | D | ✓✕✓ correctKey D. 1.0 / 1.0 |
| Q5. Which cache policy keeps hot keys under a memory ceiling? | B | ✓✕✕ wrongKey A (bounded LRU). 0 / 1.0 |
| Q6. What does an idempotency-key header protect against? | A | ✓✕✓ correctKey A. 1.0 / 1.0 |
Marking a key by hand is not judgment. It is transcription with a chance of a slip.
Auto-marked against the same key, on the same answers, the moment it was submitted.
Act 3 · Station 03 · what you are left holding
The pile becomes a shortlist
| Candidate | Q2 score | MCQ | Note | Decision |
|---|---|---|---|---|
| Rhea Kulkarni | 7? 6.5 | 10/12 | clean code, something off w/ limit | ? |
| Daniyal Ansari | 8 | 11/12 | strong. gut: yes | advance |
| Meera Subramanian | 6 | ran out of time to check tests | ? | |
| Tobias Lund | ||||
| …33 more |
Every empty cell is a candidate still waiting, and a hiring manager still asking.
- 01Daniyal Ansari4/4 tests · MCQ 11/12 · bounded LRU with explicit eviction · wrote a boundary test himself88
- 02Anika Bose4/4 tests · MCQ 10/12 · correct at the limit boundary · naming inconsistent in two places84
- 03Kwame Mensah4/4 tests · MCQ 9/12 · O(1) eviction · no note on memory growth79
- 04Rhea Kulkarni3/4 tests · MCQ 10/12 · off-by-one at line 17 · strongest readability in the batch74
- 05Priyal Shah3/4 tests · MCQ 10/12 · window expiry handled lazily · two keys leaked across tenants72
The agent scored, evidenced and ranked. It does not advance anybody. Rhea sits at 4 with one character wrong and the cleanest code in the batch. A machine will never decide whether that is a hire.
Humans keep the final call. Nothing moves until you decide.
By hand you are left with a half-filled tracker and three different scoring moods in it.
The agent hands you five scorecards to read instead of thirty-eight to grade. Same rubric on every one.
What AI Assessments does
AI Assessments gives every candidate a role-specific test, from coding challenges to MCQs, written answers, file uploads and whiteboard assessments, and grades it the moment it is submitted. Instead of a raw pass or fail, you get a scored breakdown per skill with the evidence behind each mark, so a shortlist reflects genuine capability.
Sessions are proctored, and the integrity signals are reported rather than enforced. The agent does the marking; a human keeps the final call on who moves forward.
Role-specific and coding tests
Coding, MCQ, written-answer, file-upload and whiteboard formats, matched to the role, not a generic quiz.
Auto-graded on submit
Results resolve the moment a candidate submits, with no manual marking queue.
Evidence per skill
Each score points at what the candidate did, including specific lines flagged in code.
Proctoring as signals
Integrity signals are reported, never auto-rejecting anyone. A human reads the flag and decides.
Metered per assessment
You pay per assessment, so cost tracks with volume.
Pairs with interviews
Assessment scores sit beside interview results under one candidate record in the Inbuilt ATS.
What teams say
Sample quotes shown for layout review. Real customer stories will replace these.
Same test, graded the same way every time, with the failing line pointed out. Debriefs got a lot shorter.
We moved from take home guesswork to a scored breakdown per skill. The shortlist finally reflects who can do the work.
Proctoring as a signal, not a verdict, was the difference. My team reads the flag and decides.
Frequently asked questions
What kinds of tests can I run?
AI Assessments supports role-specific coding tests, MCQs, text-based questions, file-upload tasks and whiteboard assessments, plus psychometric tests and soft-skills evaluation. Each is configurable per role, covering technical and non-technical positions.
How is grading done, and how fast?
Tests are auto-graded the moment a candidate submits. You get a per-skill breakdown with the evidence behind each score, including specific lines flagged in a code submission, so results are both fast and defensible. Once the scoring is done, a recruiter can review the results and push candidates across stages.
How does proctoring work?
Sessions are proctored and the integrity signals are reported, never enforced. Nothing auto-rejects a candidate. A human reviews any flag alongside the submission and decides what it meant.
How is it priced?
Assessments are metered per assessment, so your cost tracks directly with volume. You sign up, get free credits on us, and can run real assessments and see ranked reports before you commit.
Does it work with interviews and our ATS?
Yes. Assessment scores sit alongside interview results under one candidate record and flow into the Inbuilt AI-upgraded ATS, and IntervueBox integrates with major ATS and HRMS platforms.
Works with the rest of the platform
Grade on capability, at any volume.
Assemble a role-specific test, send it, and get ranked, evidence-backed scores the moment candidates submit. You keep the final call on who advances.