The CodeSignal OA
- Know the Ramp OA format: one CodeSignal problem, about four locked levels, about 90 minutes, scored out of 600
- Know what a perfect score gets you, and how long the link stays open
- Build the banking system and the in-memory database level by level in runnable Python
- Spend the clock where the time goes, on levels three and four
Apply to Ramp cold as a software engineer and the online assessment often lands within minutes, before anyone has read your resume. It's a CodeSignal build. You get one problem that grows over about four levels in roughly 90 minutes, scored out of 600, and each level stays locked until every test on the level below passes. Referred and recruiter-sourced candidates more often start with a recruiter call, and some of them skip the OA.
The 2026 reports name the domains: an in-memory database, a banking system, a task manager, a file system, a recipe manager and, in October, container operations. Candidates in early 2026 described a new version of the OA, so expect the prompt to rotate. The shape holds. You build a stateful class, each level adds rules to it, and the old methods have to keep passing.
How the Ramp OA works
- Structure: 1 problem, about 4 levels. One class that grows. A level opens only after every test below it passes, so a bug at level two blocks the rest.
- Time and score: About 90 min, out of 600. Your score comes back when you submit. Several candidates found the clock harder than the problems.
- Window: Often about 3 days. Reports from May to July 2026 say the link expired after three days. A January 2026 candidate had about two weeks.
- Platform: Usually CodeSignal. A March 2026 frontend candidate got a 90-minute CoderPad test instead, and some 2025 intern and AI roles used HackerRank.
CodeSignal calls this format its Industry Coding Framework, and candidates have used that name for Ramp's test from 2023 through late 2025. Each level ships with tests, some of them hidden. That lock makes the format unforgiving. A 2024 candidate failed a single test on level three and never saw level four. So a level-one bug costs you three levels of points.
Plan around the short window. Three days is tight with a full-time job, and a May 2026 candidate with more than seven years of experience was still sent the test. Referred seniors sometimes skip it. If you're applying cold, practice the format before you hit submit, because the clock on the link may start that afternoon.
What happens after you submit
The score shows up right away. The decision usually comes by automated email one to two weeks later, and rejections rarely say why. Some flows, mostly intern and cold applications, add a recorded one-way video on Hireflix after the OA. Pass and you get a recruiter call, then a one-hour live screen that is very often the URL maze, where you crawl an API over HTTP.
Other versions of the OA
Frontend. Frontend applicants get a React test instead, about an hour long, and the build reported most in 2026 is a Wordle-style game. You score each letter, handle repeated letters and cap the guesses. A June 2026 candidate who got it advised being very quick, and one 2026 report gave it 45 minutes. Some versions pair the build with a short algorithm question. It's worked in Frontend Builds.
More progressive builds. The file system, task manager and recipe manager builds from 2026 are worked the same way as the two below, in More CodeSignal Builds.
How to work a four-level build
Read level one and write the plainest class that passes it. Don't design for level four before you've seen it. You'll guess wrong, and the guess eats minutes you need later. Keep state in plain dicts keyed by id, one per fact you track, so the next level adds a dict or a branch. Run the whole test file after every level, since a new rule often breaks an old method.
Python is the common choice, and any language you know well works. Write it fast, with names you can read at minute 80. When a level's tests go green, move on. Polish never earns points while level four is still locked.
The banking and database builds below are hash maps underneath. The level that hurts is the one that adds time: cashback that lands a day later, a field that expires, a backup restored with its clock still running. Pass the timestamp into every method from level one and decide early how each call learns what time it is.
Asked at Ramp
The banking system
The banking build has the longest record of any Ramp OA problem, reported from 2023 into 2026. A March 2026 candidate described "a financial system with 4 diff questions" on CoderPad. A September 2025 candidate summed up the start as initializing users, updating balances, withdrawing and sending money, with the third and fourth steps progressively harder than the first two.
Later levels add rankings like top spenders, payments with cashback or scheduled transfers, and account merging. The class below builds something like all four. Which feature sits at which level varies, so practice the shape.
Banking System With Cashback and Merges
Problem
BankingSystem. Every method takes a timestamp in milliseconds first, and timestamps only go up. Level 1. create_account(timestamp, account_id) returns true, or false if the id exists. deposit(timestamp, account_id, amount) returns the new balance. transfer(timestamp, source_id, target_id, amount) returns the source's new balance. Both return null for an unknown account. A transfer also fails from an account to itself or without enough money.Level 2.
top_spenders(timestamp, n) ranks accounts by total money sent out, highest first, with ties broken by id. It returns up to n strings like "acc1(300)". An account that sent nothing counts as 0.Level 3.
pay(timestamp, account_id, amount) withdraws money and returns an id, "payment1", "payment2" and so on, or null if the account is unknown or short. Payments count as spending. A 2% cashback, rounded down, arrives 86400000 ms later. get_payment_status(timestamp, account_id, payment_id) returns "IN_PROGRESS" or "CASHBACK_RECEIVED", or null if that account doesn't own the payment.Level 4.
merge_accounts(timestamp, account_id_1, account_id_2) folds the second account into the first and deletes it. The first takes over its balance, its spending and its payments, including cashback still on the way. It returns false for the same id twice or an unknown account.- BankingSystem()
- create_account(1, "acc1")→ true
- create_account(2, "acc2")→ true
- deposit(3, "acc1", 500)→ 500
- transfer(4, "acc1", "acc2", 200)→ 300
- transfer(5, "acc2", "acc1", 300)
acc2 holds only 200, so the last transfer fails and returns null.
Keep balances and spending totals in two dicts keyed by account id. Cashback goes on a min-heap keyed by due time, and every method starts by crediting whatever is due.
Banking System With Cashback and Merges
Click "Run" to execute your code
The in-memory database
The in-memory database has come up from late 2023 to July 2026. A June 2026 applied AI engineer had to build one as a single progressive question in four levels over 90 minutes. Another June 2026 candidate summed theirs up as hashmaps and snapshotting. A July 2026 report describes the same build, with each round adding more to the database.
The levels tend to run in this order. Level one sets, gets and deletes fields inside records, and level two scans a record's fields. Level three adds timestamps and a time to live (TTL), and level four takes backups and restores them. Older versions swap pieces, like a 2025 intern's HackerRank build with ownership, so the class below fixes one version of the shape.
In-Memory Database With TTL and Backups
Problem
InMemoryDB of records. Each record key holds fields with string values. Every method takes a timestamp first, and timestamps only go up. Level 1. set(timestamp, key, field, value) writes a field. get(timestamp, key, field) returns its value, or null. delete(timestamp, key, field) removes it and returns true, or false if it wasn't there.Level 2.
scan(timestamp, key) returns a record's fields as strings like "city(nyc)", sorted by field name. scan_by_prefix(timestamp, key, prefix) keeps only the fields that start with prefix. A missing record gives an empty list.Level 3.
set_with_ttl(timestamp, key, field, value, ttl) writes a field that lives from timestamp until just before timestamp + ttl. Every read and delete treats an expired field as missing. A plain set on the same field clears its TTL.Level 4.
backup(timestamp) saves the database and returns how many records hold at least one live field. restore(timestamp, timestamp_to_restore) brings back the latest backup taken at or before timestamp_to_restore. A restored field with a TTL keeps the time it had left at backup, counted from the restore. A suitable backup always exists.- InMemoryDB()
- set(1, "user1", "name", "ana")
- get(2, "user1", "name")→ "ana"
- delete(3, "user1", "name")→ true
- delete(4, "user1", "name")→ false
The second delete finds nothing left to remove.
Store each field as [value, expires_at], with expires_at None for no TTL. Filter out expired fields on every read. A backup stores time left instead of the absolute expiry.
In-Memory Database With TTL and Backups
Click "Run" to execute your code
How to practice for the Ramp OA
Build each class from a blank file on a 90-minute timer, one level at a time, and run every test after each level. Do the banking system and the in-memory database twice, a few days apart, since they have the most reports. The second run shows whether level one now takes ten minutes. If it does, you have time for the levels that decide the score.
Common questions
How hard is the Ramp OA?
Each level is easy to medium on its own, and the clock is the hard part. A June 2026 applied AI candidate found the first two levels easy and said the third could take time. A 2025 candidate called the questions fairly easy but impossible to finish in an hour and a half, and another scored 540 of 600 because time ran out. Expect levels three and four to take most of your 90 minutes.
Is the Ramp CodeSignal OA proctored?
It can be. In September 2025 one candidate could choose between a proctored and an unproctored sitting, and an October 2025 candidate described a proctored run. No 2026 report settles it, so read your invite. Either way, no report says AI help is allowed on the OA. Ramp's AI-assisted coding comes later, in a dedicated onsite round that many 2026 loops include.
What score do you need on the Ramp CodeSignal?
Ramp publishes no cutoff, and no score guarantees a call. Candidates were rejected at 570 and 590 in 2024, at 540 in 2025, and at a perfect 600 as recently as June 2026. Aim to clear all four levels anyway. A lower score can only hurt, and the full score at least leaves the decision to your resume.
Does the Ramp OA use LeetCode questions?
Mostly no. The default for backend and generalist roles is the progressive build, which tests clean class design under time pressure. Some roles get classic problems instead. A December 2025 iOS candidate had a 45-minute string problem, and an October 2025 AI product engineer got two LeetCode-style problems on HackerRank.
Key Takeaways
- ✓The Ramp OA is one CodeSignal problem in about four locked levels, roughly 90 minutes, scored out of 600.
- ✓Full scores still end in automated rejections, because the resume is screened after the test.
- ✓Keep state in plain dicts and pass the timestamp everywhere, so the time-based levels stay small edits.
- ✓Banking and the in-memory database lead the reports. Practice both from a blank file on a timer.
Ramp interview guideKnow what Ramp asks before you walk in.
29 coding · 3 system design · 18 behavioral questions
Get their questions first.
InterviewLegend is independent and not affiliated with or endorsed by any company it covers.