Big tech is moving away from blank-slate algorithm puzzles. The new screens hand you a broken codebase and an AI assistant, then watch how you use it. This is practice for that interview, and you get scored on how you worked, not just whether the tests went green.
💻 The practice workspace is built for desktop. Create your account now and practice from your computer.
Three stages. No hand-holding between them.
Select from multiple real interview formats across Easy, Medium, and Hard. Each is a broken implementation waiting to be fixed.
A live test runner shows what's failing. An AI coach answers questions and escalates hints, but never writes the fix, so the work stays yours the way it would in a real screen.
Six axes covering how you solved it, how you used the AI, and how you verified the result. Backed by a full timeline of every decision you made.
Six axes, measured from what you actually did. Every edit, test run and question is timestamped, so the score reflects your process, not a guess at it.
Diagnosis. Whether you understood the failure before you started changing code.
Recovery. When a test went red, whether you adapted or kept trying the same thing.
Independence. Whether you attempted your own fix before escalating to the coach.
Precision. The quality of what you asked. Specific questions score above "what is wrong".
Test Ownership. Whether you drove the test runner as your feedback loop or guessed.
Verification. Whether you reviewed AI output before applying it, or rubber-stamped it.
Two of these are computed directly rather than judged: applying AI code in under twelve seconds without reading it counts against Verification, and reveal the answer key and Independence is capped no matter what else you did.
Not LeetCode. Not flashcards. The interview companies are actually running now.
The format big tech is moving to: inherit code that is already broken, work out why, and fix it without rewriting.
Hints escalate from broad area to specific line, and the coach will not confirm a guess. How you lean on it is part of the score.
Passing the tests is the floor. The score reflects how you got there: what you tried first, what you asked, what you checked.
EventEmitter (Meta), LRU Cache (Amazon), Prefix Trie (Google), Rate Limiter, Transaction Engine, and 6 more.
Adds realistic interviewer pressure: scope creep, over-engineering bait, correctness doubts. Practice staying focused.
Instant feedback on every edit. Hidden tests reveal edge cases once you complete each part.
Each one is a real broken implementation. A sample:
Three bugs: listeners fire in reverse order, off() wipes the whole event, once() never removes itself.
Sentinel nodes never linked, eviction hits the wrong end, and get/put skip MRU promotion entirely.
Drops the last field on every row, strips leading whitespace, and enters quote mode at the wrong character.
search() returns True for prefixes, get_words_with_prefix() builds results in reverse, case handling is inconsistent.
+ 7 more: Rate Limiter, Feature Flags, Log Query Filter, Dependency Resolver, Merge Intervals, and others.
Practice unlimited scenarios for free. Pay only for AI coaching.
No credit card required
Not a subscription
Free account. No credit card. First session in under a minute.
Create free account →