What the round looks like
The screen is an IDE in the browser with a chat panel, and the project is already there: a front end, a back end, a way to build and run both. The interviewer does not hand you a function to write. They describe something the application does wrong, and the task is to find where and fix it. In the report, the application moved money between accounts and the transfers went wrong; the question bank's summary of the format puts it as transfers that sometimes produce the wrong final balance. The interviewer may ask for a small new feature on the same code if the fix lands early.
Two things the candidate wanted the next person to know. The built-in assistant was not good, and they used it only to narrow down which function might be responsible. And the platform costs time if you have not seen it: building, running, finding output, rerunning the scenario.
Why this format grades differently
A debug round on a running application is not a coding round with a bug in it. The failing scenario is a symptom; the work is the trail from the symptom to the invariant that was broken, and the fix is a few lines at the end. The question bank's summary adds that the prompt transcript is graded and that vague prompts and accepting the first suggestion are marked down, which fits the candidate's experience of an assistant that was useful as a locator and not as a fixer.
Our read: say your hypothesis before you ask the assistant anything. Which handler, which invariant, what you expect to see in the data. Then use the assistant to find the function and read it yourself. A fix you cannot explain against the symptom is the thing this round is designed to catch.
What is confirmed, and what is not
- Confirmed: a new AI-assisted format in the virtual onsite; a HackerRank IDE with a built-in assistant; a preloaded, runnable application with a front end and a back end; a stated failing scenario about transfers to diagnose and fix. The format appears in a first-hand report and in the question bank's summary of it.
- Reported once: the Cursor-like interface; the assistant's weakness; the time lost to the platform; the small feature as a possible follow-up.
- Not known: the languages of the preloaded application, how the round is scored against the classic coding rounds in the same loop, and whether the scenario varies by team.
How we would prepare
- Practise debugging from a symptom, not from a failing line. Take a small service you know, break an invariant two layers away from where it shows, and time yourself finding it.
- Learn a browser IDE's rhythm in advance. Where the run button is, where output goes, how to rerun one request. Ten minutes on any online IDE with a build step pays back the ten the reported candidate lost.
- Use the assistant as a map. Ask it which function handles the scenario and what it touches. Do not ask it to fix the bug until you can say what the bug is.
- Check the contract, not the symptom. A transfer that ends with the wrong balance can be a missing save, the wrong row, a check in the wrong order. Read what the code promises before you decide which one it is.
Nyrion's Uber section starts with Ledger Transfers, built from this report. A small ledger service with a contract and planted defects: the first ticket says transfers end with the wrong balances, the second that retried requests are posted twice, and the third asks for a daily limit as the small feature on top. The format, the transfer domain and the symptom are what the candidate describes; the codebase, each planted bug and the feature are ours, and it runs in plain Python rather than a front end and a back end.
How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.
Frequently asked, briefly answered
What is the format?
A HackerRank IDE with an interface a candidate compared to Cursor and a built-in AI assistant you may use. A small application with a front end and a back end is preloaded; the platform builds and runs it in the browser. The interviewer states a scenario that goes wrong and you find and fix it. A small feature on top of the same code may follow.
What was the bug about?
Money transfers between accounts. The candidate's words: transferring money goes wrong and you have to debug where the problem is. The question-bank summary of the format puts it as transfers that sometimes produce the wrong final balance.
How good is the built-in assistant?
Not very, in the one first-hand account: the candidate used it only to find which function might hold the problem. Plan to read the code yourself and use the assistant as a locator, not a fixer.
What costs candidates time?
The interface. The reported candidate wrote that not knowing the platform eats a lot of time. Spend the first minutes learning where the app runs, where the logs are and how to rerun a scenario before you touch the code.
Are the prompts graded?
The question-bank summary says the prompt transcript is part of the evaluation and that vague prompts and accepting the first suggestion are marked down. We have not seen that confirmed first-hand.
Practice problems tagged for Uber
Browse the catalog and open the Uber section. No login required to look; sign in when you want to solve one.