Giving back to the forum. I interviewed for a Staff SWE role at Cash App — two technical screens back to back on the same day (one coding, one called an "AI Native" screen), then a separate HM phone screen later. All three were remote.
Coding: 60 min, CoderPad
AI Native: 30 min
HM phone screen: 30 min
It's been almost two weeks since I finished and I haven't heard anything — I'm guessing the position got filled — so let me share the questions first.
Interview Questions
One: Coding (60 min, CoderPad)
First round, coding, done on CoderPad. The question was pretty close to their real business — don't go in with a LeetCode mindset.
You're asked to implement an in-memory BIN database. A BIN is the first few digits of a card number (credit / debit / prepaid) that you can use to figure out things like the card's brand and type. It starts out simple, then they layer on requirements one at a time:
Base version: two methods, store and lookup. The key is a 6-digit BIN, the value stores brand and type. When you store, both brand and type are required; assume all input is valid so no validation is needed. Writing the same key again is last-write-wins.
Q2: Add a source / trust level. Every store call now carries a trust level, and lookup returns the entry with the highest trust for that BIN. If trust levels tie, it's last-write-wins again. Duplicate brand+type combos are allowed. How you model trust is up to you.
Q3: Switch to partial updates. store no longer requires both brand and type — you might get just one (but at least one, and source is always present). lookup now has to merge per attribute: each field takes the value from whichever source gave that field and has the highest trust. A higher-trust source that didn't provide a given field can't overwrite the existing value with null.
Q4: Prefixes can now be any length, not just 6 digits, for both store and lookup. lookup might hand you a full card number, and you need to shrink from the longest prefix down until you've filled in all the needed fields. Merge gets one more rule: look at the longest (most specific) prefix first, and only compare trust within the same prefix length. By this point it was basically a data structures conversation — I brought up a prefix tree / trie, and we talked through the complexity tradeoffs of lookup versus update.
Time was mostly enough — the last question leaned more toward discussion and I wasn't asked to fully implement it.
Two: AI Native (30 min)
The second round was a screen called "AI Native" — very open-ended, meant to see your actual day-to-day workflow and how you use AI. No coding, no system design. You pick a recent project first, and they follow up from there. Roughly these questions:
- Give a high-level description of a recent project you worked on.
- How do you approach engineering design? What does the design phase look like?
- How do you discover and align requirements across different teams — is that info already there, or do you have to dig it out yourself?
- Once design is aligned and approved, what does implementation look like?
- Walk me through how you use AI to break down the problem space — what do you want from the agent, and how do you verify the output is correct?
- What's changed in your workflow over the past year that works really well and you've come to rely on?
- After launch, what do maintenance and iteration look like? What do you focus on post-launch?
Overall it felt pretty conversational — when answering, it worked fine to naturally bring up things like skills, knowledge bases, being data-driven, and so on.
Three: HM phone screen (30 min)
The HM phone screen was 30 minutes, with the banking / money org team. It was mostly digging deep into one project from my resume, then in the second half shifted to tech debt in legacy systems and the regulated nature of the finance space. Roughly these questions:
- Walk me through a recent high-leverage project you led — the project, its business impact, and what you did yourself vs. handed off to teammates.
- How did you decide on the tech stack / architecture?
- What did you hit on latency, performance, and reliability?
- With all those parallel services and the crawler running, how did you manage cost?
- How do you handle tech debt in decade-old legacy services with inconsistent standards? Give a concrete example.
- Money-movement systems fail differently from normal products (silent correctness bugs, regulatory visibility) — what domain strengths would you bring, and what would you need to learn?
- How do you use AI to boost your day-to-day efficiency?
- Which AI coding tools do you use? Tried any of the newer models?
The interviewer was friendly and it was a pretty smooth conversation.
Discussion
Loading comments…