The technical interview that asks the second question.

Publish one job and share one link. You read a report that explains every score.

Project round

Your résumé says you rebuilt the checkout pricing service. What did you change, and why?

Pricing hit three services on every request, so I put a read-through cache in front of them. It cut p95 latency by about 40 percent.

Follow-upHow did you measure that 40 percent?

We compared a week of p95 on the checkout dashboard before and after the rollout. Honestly, we didn't control for the traffic dip that same week.

A sample exchange. Questions are read aloud and stay on screen; the candidate speaks or types, and edits the transcript before it is scored.

It follows the answer, not a script.

The project and fundamentals rounds are a conversation. The interviewer decides what to ask next from what the candidate just said, one question at a time. Scores stay off the screen until the report.

A vague answer gets “walk me through that.”

Hand-waving doesn't earn a score. The interviewer asks for the specific case, the way a person across the table would.

Project round

How did the payment queue handle failures?

We had retries and stuff. It was pretty robust.

Follow-upWalk me through what happened when a payment failed twice in a row.

A claim gets checked.

Numbers, ownership, and outcomes on a résumé are asked about directly. A strong answer earns a harder question, not a compliment.

Project round

What was your part in the move to Postgres?

I led the migration off Mongo for the orders data.

Follow-upWhat did you move first, and how did you know it was safe to switch reads over?

What they mention gets picked up.

Something relevant said in passing becomes the next question, so the fundamentals round is about the tools this candidate actually uses.

Fundamentals

Which part of the ingestion pipeline did you build?

The worker that parses vendor files. I wrote it in Java with a small plugin system.

Picked up from their answerYou mentioned Java and plugins. When would you pick composition over inheritance there?

It knows when to stop.

Not knowing is allowed. There is at most one simpler angle, then the interviewer moves on. It never repeats a question or drills a lost candidate.

Fundamentals

I'm honestly not sure how the kernel schedules threads.

Follow-upThat's fine. What's the difference between a process and a thread?

Not sure, to be honest.

New topicLet's switch. Walk me through how a request reaches your API.

The tests decide. The agent explains.

The coding round is a real editor and a real judge. Code runs in a sandbox against tests written for your job, and the review adds likely complexity, labelled as an estimate.

Longest Substring Without Repeats31:08

Given a string s, print the length of the longest substring that contains no repeated characters.

Example 1

in abcabcbb

out 3

Example 2

in bbbbb

out 1

Starter code already reads stdin. Hidden tests run on submit.

1import sys
2
3def longest_unique(s: str) -> int:
4 last = {}
5 start = best = 0
6 for i, ch in enumerate(s):
7 if last.get(ch, -1) >= start:
8 start = last[ch] + 1
9 last[ch] = i
10 best = max(best, i - start + 1)
11 return best
12
13print(longest_unique(sys.stdin.readline().strip()))
Run samples
  • abcabcbb→ 338 ms
  • bbbbb→ 135 ms
  • pwwkew→ 336 ms
Accepted11 of 11 hidden tests passedLikely O(n) time against an O(n) target · estimate

Written for your job

The agent writes the problem, starter code, and tests from the job description. Expected outputs come from running a reference solution, which is not stored.

Hidden tests stay hidden

Samples run as often as the candidate likes. Submit runs the hidden tests and ends the round. Hidden inputs are never shown.

Nothing is lost

Drafts save per problem and language. The timer lives on the server, so a refresh doesn't reset it. When time runs out, the editor is submitted.

13 languages

Python, JavaScript, TypeScript, Java, C++, Go, C, C#, Rust, Kotlin, Ruby, PHP, Swift.

A report that explains every score.

Every skill discussed gets a rating and the evidence for it, taken from what the candidate said. A round score averages those ratings, so a shaky first answer they recover from is not counted against them twice.

  • Round scores, and an overall score that averages the rounds completed
  • A rating for each skill discussed, with the evidence behind it
  • The submitted code, its test results, and its likely complexity
  • Questions for a follow-up interview, aimed at what this one could not confirm
  • Every camera and browser observation, listed for review

Sample report · illustrative candidate

Priya Raman

Backend Engineer, Payments · Finished

78

Overall

Coding
92
Projects
74
Fundamentals
68

Project round

You designed the caching change well and were honest about how it was measured. Testing before release is the gap for this role.

  • Caching and service designStrong

    You explained why pricing was the hot path and put one read-through cache in front of three services. You raised invalidation on price changes without being asked.

    88
  • Performance measurementSolid

    You said the cache cut p95 latency by about 40 percent. Asked how you measured it, you compared a week of dashboard p95 before and after, and noted the traffic dip that week yourself. A strong answer would have compared like-for-like traffic or used a load test.

    76
  • TestingDeveloping

    You described checking the dashboard after release, but no tests before it. A strong answer would cover how stale prices are caught.

    58

For a follow-up interview

  • How would you verify the latency gain under the same traffic?
  • How would a stale price be caught before a customer saw it?

Coding review

11 of 11 hidden tests passed. Likely O(n) time against an O(n) target, labelled as an estimate.

3 warnings in the monitoring log

It records what happened. A person decides what it means.

The camera and browser note a missing face, a second person, a phone, leaving the tab or full screen, and eyes held off the screen. Nothing here decides that someone cheated.

  • One continuous condition is one warning, however long it lasts.
  • A glance at the keyboard waits several seconds before it counts.
  • Speaking an answer is expected, so mouth movement never adds a warning.
  • Five warnings lock the round. Each one tells the candidate what was observed.
Monitoring logSample session
  1. 00:03:12Tried to leave full screen, returned in 2 sWarning 1 of 5
  2. 00:11:47Switched tabs for 6 sWarning 2 of 5
  3. 00:18:30Looked down at the keyboardNot counted
  4. 00:24:05Eyes held off the screen for 5 sWarning 3 of 5
  5. 00:31:52Mouth movement during a spoken answerKept for review

From a job description to a scoreboard.

  1. Publish the job

    Paste the job description and what you care about. The agent writes the coding problem and its tests, and plans the spoken rounds. It takes a minute or two.

    Backend Engineer, Payments

    Go or Java, Postgres, queues. Owns production incidents end to end.

    Focus: data modelling, debugging under pressure

  2. Share one link

    Candidates sign in, add their résumé, and start. Spoken questions are written from that résumé and your job, so no two conversations are the same.

    /i/k7q2-backend-payments
  3. Read the scoreboard

    Name, status, per-round scores, and the overall score for everyone who attempted it, each linked to their written report.

    Backend Engineer, PaymentsSample
    CandidateCodeOverall

    Priya Raman

    Finished

    9278

    Daniel Okafor

    Finished

    7680

    Mei Tanaka

    In progress

    88—

Send the interview. Read what happened.

Publishing takes a minute or two. Interviewer accounts need an access code; candidates only need the link.