What the round looks like
You arrive with your tools set up, which the stipend is for. The interviewer shares a repository and a brief. In the quoted version, the brief explains the problem first: finding the right transaction, clicking into it and uploading one receipt at a time is tedious. Then the new capability: the backend team shipped an OCR service that reads a receipt image, extracts the amount, merchant and date, and suggests which transaction it matches, in a few seconds. Then the ask, as four needs: upload receipts, show processing status, display the OCR results, let the user confirm or change the suggested match. Finish and the interviewer offers bonus items.
The second report is shorter and matches: one AI live coding round in the onsite, a stipend for tools, a repository, a matching system to build, language unrestricted, bonuses if there is time.
What it measures
The candidate who wrote the longer report also wrote their read of the grading: beyond whether you can build it, the interviewers probably watch how you break the requirement down and how you verify what the assistant generated. That matches the brief's shape. Four needs are listed, and they are not equal: the processing state is the one that forces a design decision, because a few seconds of OCR with no callback means the client polls or the server pushes, and either way the user has to see where their receipt is.
Our read: the round rewards a candidate who turns the brief into a short ordered list before the first prompt, builds the smallest vertical slice that shows a receipt going from uploaded to extracted to matched, and reads the assistant's code against the brief at each step. The bonus items are there to see what you reach for once the core is honest.
What is confirmed, and what is not
- Confirmed: an AI live coding round in the onsite; a stipend to set up any AI tools; an existing repository; a feature in any language; bonus items when the main part is done. Each appears in both reports.
- Reported once: the brief's exact wording (receipt upload, the OCR service, the four needs); the view that decomposition and verification are what is graded; a prompt-history file handed in with the work.
- Not known: the repository's language and size, what the bonus items are, how the round is scored against the rest of the onsite, and whether the brief changes between candidates.
How we would prepare
- Turn a brief into an ordered list in two minutes. Take any product request and write the four or five things it needs, in the order you would build them, with the one that forces a design decision marked.
- Build a vertical slice first. One receipt, uploaded, processing, extracted, matched. Everything else is a second pass; the interviewer sees working software early.
- Design the asynchronous part on purpose. A job that takes a few seconds and never calls back means the status is brought up to date on read or pushed. Say which, and why.
- Read the generated code against the brief, line by line.Does confirm release the old transaction? Does a failed extraction have a visible state? That reading is the verification the candidates believe is graded.
- Keep a list of bonus items you would want. Batch upload, a queue of receipts the service could not place, undo. Being ready to choose is part of finishing early.
Nyrion's Ramp section starts with Receipt Matching, built from these reports. A receipts service over a fake OCR that takes a few seconds and never calls back: the upload flow with visible processing state, then the suggested transaction with confirm, change and unmatch, then the bonus stage, a batch upload and a review queue of receipts the service could not place. The brief, the OCR fields, the few seconds and the confirm-or-change step are what candidates describe; the backend shape, the matching rule, the state machine and the bonus content are ours, and it is the service behind the UI rather than the UI.
How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.
Frequently asked, briefly answered
What is Ramp's AI live coding round?
One round of the onsite. Ramp pays a fifty-dollar stipend to set up whatever AI tools you want, hands you an existing repository, and asks for a feature in any language. If the main part is done with time left, you pick bonus items. Two candidate reports from July 2026 describe it the same way.
What was the feature?
A better receipt upload and matching experience. In the brief, as one candidate quoted it: the current flow is tedious, users find the right transaction, click in and upload one receipt at a time; the backend team just shipped an OCR service that extracts amount, merchant and date and suggests a matching transaction, taking a few seconds; build an upload flow that shows processing status, displays the OCR results and lets the user confirm or change the suggested match.
What is graded?
Beyond whether it works: how you break the requirement down and how you verify what the assistant produced. One candidate's read, after the round, was that more and more companies run this format and that decomposition and verification are what the interviewers are watching.
Which language and tools?
Any. The repository is given; the language of your feature is yours; the tools are whatever the stipend bought. One earlier report described handing in a prompt-history file with the work.
What are the bonus items?
Reported only as items you pick once the main part is done. We do not know their content. Our problem treats batch upload and a review queue for receipts the service could not place as the bonus stage; that part is ours.
Practice problems tagged for Ramp
Browse the catalog and open the Ramp section. No login required to look; sign in when you want to solve one.