A project deep dive interview scores the decisions you made on one real project. A smaller project can beat a bigger one. Shopify’s guide to its 60-minute Technical Deep Dive lists what interviewers want to hear. Your role and the architecture are on it. So are the tradeoffs, the alternatives, the failures and what you’d change. Ashby runs a similar round with its VP of Engineering. Its guide names projects it’s much less interested in. Migrations, cross-team coordination and work you left before release are on that list. So pick the project where you made the most decisions yourself. Prefer one that shipped, where you saw the result. Then rebuild every decision before the interview. Note the options you had, why you chose, what it cost and what you’d change. In the room, give a two-minute map and let them interrupt. Most of the time goes to their follow-up questions.
The project deep dive interview scores decisions you owned
Two company guides describe this round in detail. Shopify’s Engineering blog post from July 2022 describes a 60-minute conversation with engineers. They want specific technical detail. The topics start with your approach, business impact and architecture. They continue through technical decisions, alternatives, failures and lessons. Shopify also asks what you would do differently.
Ashby’s candidate guide was last updated in July 2025. Ashby sells recruiting software. This page describes its own hiring. The walkthrough happens twice. The first is about 10 to 20 minutes in an intro call with an engineering leader. The second is a 30 or 45 minute final round with the VP of Engineering and another leader. Ashby says the length depends on seniority. Its reason for the round is plain. Most interviews simulate work under artificial time limits. A past project shows how you actually gathered information, made decisions and handled setbacks.
Both guides push the same way. The interviewer wants to know which calls were yours and how you made them. A big project with no decisions you can defend scores poorly. A smaller project full of your own tradeoffs scores well. Ashby lists the signals of a strong pick. You led it and made the critical calls. You remember specifics such as data model fields and milestone dates. It shipped, so you saw what happened.
Pick the project with the most decisions you made yourself
Take an Android engineer we’ll call Tariq, with two candidate projects. The first is the team’s XML-to-Compose migration across 40 screens. Tariq ran the rollout plan and reviewed most of the pull requests. It’s the most visible work on the resume. The second is smaller. Field inspectors were losing reports when they filed from basements with no signal. Tariq designed and built an offline outbox for the inspection app. It shipped. Lost reports stopped.
The migration looks like the safer pick. Ashby’s guide says otherwise. It lists migrations among the projects it’s much less interested in. Its reason is that they rarely involve product or business-driven technical decisions. It also lists projects with little ambiguity that mostly needed coordination across teams. Ashby adds that process-heavy teams value that work more than it does. So a coordination project isn’t wrong everywhere. It just gives this round less to probe. Most of the migration’s hard calls came from Google’s Compose guidance and the team’s style rules. Tariq applied them carefully. Someone else made most of them. An interviewer who asks “why that approach?” hears “that’s the recommended way” several times.
The outbox is full of Tariq’s own calls. How should a queued report be stored? When should it upload? What happens if the same report reaches the server twice? Each question had real alternatives. Tariq chose between them. Shopify adds one more filter. Its guide steers you away from solo or side projects. Those often lack business tradeoffs and product context. The outbox is team work with a real user problem behind it, so it passes.
If you’re unsure, Ashby lets candidates send one or two ideas to the hiring manager first. Not every company offers that. Asking still costs nothing. Our hiring-manager screen post covers that conversation.
Rebuild a decision log before you rehearse the story
Most candidates prepare this round as a story. They practice the problem, the build and the result. Then the first follow-up asks why they used one approach over another. They can’t remember the alternative. The story stalls on its first real question.
Prepare a decision log instead. Write one entry for every call you’d expect a probe on. Each entry has four parts. Name the options you had. Say why you picked one. Say what the choice cost. Say what you’d change now. Then add the specifics Ashby says strong candidates remember. These are table and field names, rough dates for each milestone and how users reacted.
For Tariq’s outbox, the log has five entries:
- Storage. A Room table for pending reports, chosen over files on disk. Room gave a query for “what’s still pending” and survived process death. The cost was a schema migration in the next release.
- Upload trigger. WorkManager with a network constraint and exponential backoff, chosen over retrying in the screen. Android’s WorkManager docs describe both the network constraint and the backoff policy. The cost was less control over exactly when uploads ran.
- Duplicates. A UUID created on the phone and sent with every upload. The server ignored any ID it had already stored. The cost was a backend change and a second team to convince.
- Scope. A one-way outbox, not full two-way sync. Inspectors only ever created reports offline. They never edited old ones. Two-way sync would have doubled the work.
- The miss. Photos were 4 to 8 MB each. The first version queued them with the report and blew past upload timeouts. Tariq split photos into their own work requests in the second release.
Plan to talk about the miss. Shopify lists failures and lessons as topics the interviewer wants. Ashby lists challenges and lessons learned. A project with no visible mistake sounds rehearsed. It also gives the interviewer nothing to probe.
Open with a two-minute map, then let them interrupt
Use the opening to give the interviewer a map, so they can choose where to dig. Ashby’s guide says you lead and the interviewers stop you with questions. Shopify’s guide is blunter. “Remember this is a conversation not a presentation.”
So don’t plan 20 minutes of narration. Plan two minutes that name the problem, your role, the shape of the system and the result. Then stop and offer the decisions as places to go. The interviewer picks one. The decision log carries the rest.
Slides are optional at both companies. Ashby says a simple diagram is fine if it helps you explain. For an Android project, one box diagram is usually enough. Show the screen, the local table, the worker and the server.
A sample opening and three probe answers
Here is Tariq’s opening, about two minutes spoken:
“I’ll walk through an offline outbox I designed and built for our inspection app. Field inspectors file safety reports from basements and plant rooms, often with no signal. Support was getting about a dozen tickets a week about reports that vanished. They hadn’t vanished. The app sent each one once. When the request failed, nothing retried. I owned the client design and most of the build. A backend engineer owned the server change.
“The design is a one-way outbox. When an inspector taps submit, the report goes into a Room table first. It gets an ID generated on the phone. A WorkManager job uploads it once there’s a network. It retries with backoff. The server drops any ID it’s already seen, so a retry can’t create a duplicate. We shipped it in March. Lost-report tickets went to zero within a month. We did have one bad miss with photo sizes. I can get into that.
“There are three decisions I’d be glad to go deeper on. One is why I used Room over files. Another is how we stopped duplicates. The third is why I didn’t build full sync. Or I can start wherever you’d like.”
Then the follow-ups come. Here are three likely probes, each answered from the log.
Probe: “Why not retry from the ViewModel when the network comes back?”
“I considered it. The problem was process death. Inspectors switch apps a lot on site, so Android can kill our process in the background. A retry tied to the screen dies with it. WorkManager persists the job, so it runs even if the app was killed. The cost was less control over timing. Uploads can lag a few minutes after signal returns. Inspectors didn’t care about minutes. They cared about losing the report.”
Probe: “What would you change?”
“Photos. I queued them inside the same job as the report text. A batch of 6 MB photos timed out over weak signal, so the whole report kept retrying. I split photos into their own jobs in the next release. Now the report text lands first. If I did it again, I’d size the photos before writing any queue code.”
Probe: “What if this had to support editing reports offline?”
“Then the outbox design breaks. Two edits to the same report could reach the server in either order. So I’d need a conflict rule. I didn’t build that, because inspectors never edited offline. I checked that with the product manager before cutting it. If that changed, I’d start by asking the server team who wins a conflict.”
Why the opening and the answers work
The opening names the user and the cost in one sentence each. It names inspectors in basements and a dozen tickets a week. Those details cover the business impact Shopify asks about without a speech about it.
It draws the line between Tariq’s work and the backend engineer’s. Ownership is what the round is checking, so state it before anyone asks. Claiming the server change would fall apart at the first server question.
It mentions the photo miss in the opening, so Tariq raises it before the interviewer finds it. The interviewer hears that Tariq knows what went wrong and why.
It ends with three named decisions. The interviewer can follow one of them or go elsewhere. Either way, Tariq has already flagged the places where the log runs deepest.
Each probe answer follows the log entry. It names the alternative, gives the reason, then states the cost. Leave out the cost and the decision sounds like the only option. Then the interviewer learns nothing about judgment.
The third answer goes past what Tariq built. It says how the design would break and what Tariq would ask first. Shopify’s “what would you do differently” topic often turns into a question like this. A full design isn’t expected here. Showing where your own design stops is enough.
Say “that wasn’t mine” when the interviewer drills into someone else’s work
Sooner or later the interviewer asks about a part you didn’t build. For Tariq, it’s the server’s duplicate check. Was it a unique index? Did it hold up under load?
The weak move is to guess with confidence. The interviewer is technical and will probe the guess. The strong move is short. Say who owned it, what you asked them for and what you checked. “The backend engineer owned that. I asked for one rule. A repeated ID gets a success response and no new row. I tested it by sending the same report three times from a test build.” That answer shows the boundary and the part you did own.
The project should also fit your level. A senior candidate’s project should show calls that touched other teams. Tariq’s backend negotiation does that. For staff-level loops, expect the interviewer to push on scope and influence across teams. Our senior Android system design post covers how that bar shows up in a design round. The deep dive applies the same bar to work you’ve already done.
1 thought on “Project Deep Dive Interview: How to Pick a Project You Can Defend”