DoorDash AI CodeCraft Interview: 8 Experiences

Shared by janebeingjane ·
Company
DoorDash
Reports
8 interview reports
Level
Unknown
Stages
Onsite

DoorDash's AI CodeCraft interview in 2025–2026 was a build-and-run hour with your own AI tool: a refund workflow, a prediction service or a tracking page.

At a glance

  • Process: an onsite AI round, sometimes after a plain Code Craft phone screen (02, 03); task files came through HackerRank but the coding happened in the candidate's own AI editor (03, 06).
  • Asked recently: a refund workflow for software engineers (01, 03, 05, 06), a prediction service for the MLE (07), and live tracking over SSE for full-stack (08).
  • Difficulty: reports disagree on speed versus care: 02 and 07 felt speed won, while 08's rubric included testing and reviewing the AI's output.
  • Tools: candidates brought their own AI editor; 06 and 08 said a browser chat alone wasn't enough.

Coverage: Eight experienced engineers in full-time DoorDash loops: six general software engineers (one at E4, and one described as E4 in replies under their post), one machine learning engineer and one E4 full-stack engineer. Seven say they were changing jobs. Every AI CodeCraft round here was part of the onsite. Seven candidates place their interviews in October–December 2026 (one gives exact onsite dates, October 7–8); the eighth, in the U.S., posted in December 2025. The reports went up between December 8, 2025 and November 9, 2026.

The three roles got different tasks, so each has its own section:

  • Software engineers (reports 01–06): four of the six describe the same refund workflow task.
  • Machine learning engineer (report 07): a real-time prediction system built from separate local services.
  • Full-stack engineer (report 08): live Dasher tracking with Server-Sent Events on top of a Node backend.

Software engineers: the AI CodeCraft round

What came up in the software engineer AI CodeCraft rounds

The refund workflow came up in four of these six reports, from December 2025 to November 2026, and two candidates said the real test was getting several services up and running live. Follow-ups pushed toward production: duplicate calls, slow workflows, what you'd change before launch.

Topic Reports Last seen
The refund workflow task 4 of 6 Nov 2026 (report date)
Production follow-ups: duplicate calls, slowness, launch changes 3 of 6 Nov 2026 (report date)
Getting several services up and running live 2 of 6 Nov 2026 (report date)
Own AI editor, with the prompt or starter code from HackerRank and the code handed back there 2 of 6 Nov 2026 (report date)
A closing recap plus an ASCII architecture diagram 1 of 6 Nov 2026 (report date)
Questions about the chosen AI tool and model while the AI was coding 1 of 6 Nov 2026 (report date)
Prompting skill named as something DoorDash grades 1 of 6 Dec 2025 (report date)

Reports 02 and 04 say their task had been reported before but don't say which, so they count only where their own details fit. A candidate can appear in more than one row.

Questions we found in the software engineer rounds

AI Coding

Question Reports Last seen
A refund workflow engine that walks a DAG of steps (check the delay, refund, notify), with follow-ups (Build a Refund DAG with Local HTTP Services) 4: 01, 03, 05, 06 Nov 2026

Behavioral (BQ)

Question Reports Last seen
Why this AI tool and this model, and have you built skills or agents at work? (asked while the AI was writing code) (practice) 1: 04 Nov 2026 (report date)

6 detailed candidate timelines, loops, and screens

01: Refund services running live in a terminal

Report date: Nov 9, 2026 · Interview period: 2026 Q4, onsite over two days · Role: Software engineer (general), full-time · Degree listed: Bachelor's · Channel: Headhunter · Result: Rejected

Day 1: system design, then debugging

  1. The candidate later blamed these two rounds for the result.

Day 2: AI coding

  1. The task was the refund question that earlier candidates had already written about. FastPrep's practice version is Build a Refund DAG with Local HTTP Services.
  2. As the candidate saw it, the round came down to one thing: could you get several microservices running live in the terminal.
  3. The conversation felt fine. They couldn't read whether it had gone well or badly.

Day 2: hiring manager

  1. Standard behavioral questions, and a good conversation.

The rejection came about half a week after the hiring manager round. In the candidate's view, the AI round wasn't what sank it.


02: An E4 pass where speed seemed to matter more than review

Interview dates: Code Craft phone screen Sep 18; onsite Oct 7–8, with the hiring manager round moved to about Oct 20, 2026 · Report date: Nov 4, 2026 · Role: Software engineer, E4, full-time · Degree listed: Master's · Channel: Referral · Result: Passed, then team matching

Sep 18: Code Craft phone screen

  1. Named only. Nothing in the report says AI was allowed here.

Onsite round 3: AI coding

  1. Another question that had shown up in earlier reports. The candidate doesn't say which, only that the key was getting a few services running fast.
  2. Their take: DoorDash seemed to care more about quick iteration than about combing through the code the AI produced.
  3. Of all the onsite conversations, this was the one they enjoyed most, and the pressure felt fairly low.

The onsite also had system design, debugging and hiring manager rounds. The whole loop is written up in 2026 DoorDash Experienced SWE (E4) and Data Analyst Loops.


03: A refund engine on a DAG, closed with a recap and an ASCII diagram

Report date: Nov 3, 2026 · Interview period: 2026 Q4, onsite over two days · Role: Software engineer (general), full-time · Degree listed: Master's · Channel: Headhunter · Result: Rejected

Day 1, round 2: AI Code Craft (1 hour)

  1. Setup: the candidate's own laptop and own AI editor, with the screen shared. HackerRank handled the hand-off: starter code came from it and the finished code went back to it.
  2. Task: the starter code already describes a refund workflow as a DAG of nodes. The job is the engine that walks it, so orders that are late or timed out get refunded without anyone stepping in. A typical path checks the delay, issues the refund and sends a notice, and each node's result decides which edge to follow next. Practice it as Build a Refund DAG with Local HTTP Services.
  3. Follow-ups came in three directions: support partial refunds and more node types such as cancel; work out what happens when the same API call is made twice or more; and make the wait shorter for the customer when the workflow is slow.
  4. The last few minutes were a wrap-up. The interviewer asked for a summary of the problem, the design choices, the trade-offs and where the code stood, then an ASCII sketch of the architecture with components, data flow and protocols labeled.

The loop also had a Code Craft phone screen, a system design round, a debugging round and a hiring manager round. All of them are covered in 2026 DoorDash Experienced SWE Phone Screens and Onsites.


04: Small talk about models while the AI did the typing

Report date: Nov 3, 2026 · Interview period: 2026 Q4 · Role: Software engineer (general), full-time · Degree listed: Master's · Channel: Headhunter · Result: Rejected, no feedback

Onsite: AI Codecraft

  1. The task was one earlier candidates had reported. This report doesn't describe it. The candidate built everything it asked for and ran an end-to-end test.
  2. While they waited for the AI to finish writing code, the interviewer asked about AI more generally: why this tool, why this model, and whether they had written skills or agents in their current job (practice).
  3. The final question: if this system were going to production, what would you optimize?
  4. Even after finishing, the candidate wasn't sure what the round was trying to measure.

The loop also had a phone screen and system design, debugging and hiring manager rounds; 2026 DoorDash Experienced SWE Phone Screens and Onsites has the rest.


05: A one-line mention of an AI refund round

Report date: Oct 31, 2026 · Interview period: 2026 Q4 · Role: Software engineer (general), full-time · Degree listed: Master's · Channel: Applied online · Result: Rejected

Onsite (four rounds)

  1. One of the four rounds was an AI round on a refund system or workflow. That is all the report says about it. The matching practice task is Build a Refund DAG with Local HTTP Services.
  2. The other rounds it names are system design and debugging.

A one-hour phone screen came before the onsite.


06: A self-help refund workflow, where prompting was said to count

Report date: Dec 8, 2025 · Location: U.S. · Role: Software engineer · Level: Mid level, SWE II Backend (E4), according to anonymous replies under the post that read as the author's own · Result: Not stated

Onsite: AI CodeCraft (60 minutes)

  1. Setup: come with a local editor that has AI wired in, such as VS Code with Copilot, Cursor or Claude Code. A ChatGPT tab in the browser doesn't count. The prompt appears in HackerRank, the coding happens in your own IDE, and at the end everything goes back into HackerRank as one file.
  2. The story: support agents keep getting tickets like "my order came late" or "an item is missing." Build the engine that runs DoorDash's self-help menu for those common cases so support spends less time on them. FastPrep's practice version is Build a Refund DAG with Local HTTP Services.
  3. The starter code had an order data class; a mock DoorDash API that can fetch an order, read its current ETA and refund a given percentage; node types START, END, IS_LATE and FULL_REFUND; a base Node class; a Workflow that holds the nodes as a DAG, with finding the start node and loading from text left as stubs; and a WorkflowEngine whose run method (taking the workflow and an order ID string) was also a stub. One test came with it.
  4. Part 1: fill in the stubs until the provided test passes, without changing the test.
  5. Part 2: a late order gets a full refund so far, which costs too much when the Dasher is only a little late. Add partial refunds. The exact rules are left for you to settle with the interviewer; reuse the existing interface and write a few tests of your own.
  6. Part 3: support says one run takes 10 to 20 seconds. How would you track that down? The interviewer then asked the candidate to pick something in their own code to improve, and they ended up building some of those improvements, with AI allowed.
  7. There were more parts after that. The interviewer controls the pace, and nobody is expected to reach the end.
  8. The candidate says DoorDash explicitly grades prompting ability, though they don't know how. Their read is that it pays to pin down requirements first and to show the AI is being steered, not doing the thinking. They also mention having failed candidates who leaned on AI too heavily, so they seem to interview people themselves somewhere.

Machine learning engineer: the AI CodeCraft round

What came up in the MLE AI CodeCraft round

The one MLE report had no refund task. The candidate had to stand up a small real-time prediction system with separate services that call each other.

Questions we found in the MLE round

AI Coding

Question Reports Last seen
A running real-time prediction system: feature store, model registry and inference service as separate services (Build a Real-Time Prediction Service) 1: 07 Nov 2026 (report date)

1 detailed candidate timeline

07: A prediction service on a laptop, with the interviewer pushing for speed

Report date: Nov 7, 2026 · Interview period: 2026 Q4 · Role: Machine learning engineer, full-time · Degree listed: PhD · Channel: Headhunter · Result: Rejected three days after the onsite

Onsite round 1: AI craft coding

  1. The candidate expected the delayed-order refund question. They got something new.
  2. The job: build a real-time prediction system that runs on your own machine. A caller asks about one entity, such as an order, and gets a fresh prediction back. The prompt's example is a model guessing whether an order will be delivered late. One piece collects the inputs for that entity (a feature store), one decides which version of the model to use (a model registry), and one runs the model and returns the answer (an inference service). FastPrep's practice version is Build a Real-Time Prediction Service.
  3. The pieces have to run on their own, talk over real API calls and stay up, so the interviewer can watch requests flow through. Boundaries, data shapes, API design, how many models or entity types, and edge cases were all the candidate's call.
  4. No training. A stub model and fake data were fine, because the model itself wasn't being judged.
  5. The written prompt said what mattered: a system that works end to end, sensible cuts under time pressure, real speed from the AI, and parts that behave like separate systems. Neat abstractions and deep class trees mattered less.
  6. The interviewer was in a hurry. They had the candidate drop the long prompt straight into the AI for a summary and then iterate fast. The candidate read that as part of what was being tested, though they're not sure it's how anyone should use AI at work.

The other rounds were an LLM design round, an ML design round and a hiring manager round. The whole loop is report 01 in 2021–2026 DoorDash Machine Learning Engineer & ML Platform Interviews, Experienced Hires (Part 1).


Full-stack engineer: the AI CodeCraft round

What came up in the full-stack AI CodeCraft round

The full-stack task was frontend work on an existing backend: live Dasher tracking over Server-Sent Events, built in milestones. This candidate also wrote down how the round was graded.

Questions we found in the full-stack round

AI Coding

Question Reports Last seen
Live Dasher tracking with Server-Sent Events on an existing Node backend, built in milestones (Build Live Dasher Tracking over Server-Sent Events) 1: 08 Nov 2026 (report date)

1 detailed candidate timeline

08: Live tracking over SSE, graded on how the AI's work was checked

Report date: Nov 1, 2026 · Interview period: 2026 Q4 · Role: Full-stack engineer, E4, full-time · Degree listed: Master's · Result: Not stated

Onsite: AI CodeCraft

  1. Tools: a real AI coding tool such as Claude Code, Cursor or Codex. A chat assistant in a browser tab wasn't enough on its own. The screen stayed shared the whole time.
  2. Task: start from a Node backend skeleton that already exists and build a frontend feature step by step, showing a Dasher's location live through Server-Sent Events. The work followed set milestones, with the live updates coming together one milestone at a time. Practice it as Build Live Dasher Tracking over Server-Sent Events.
  3. According to the candidate, the interviewer was scoring five things: how good the prompts were, why they chose their model, engineering judgment, how they checked what the AI wrote (unit tests, Playwright), and how they organized and reviewed PRs.

The same loop had a frontend phone screen, a hiring manager chat, a system design round, an Engineering Values round and a separate CodeCraft round where AI wasn't part of it.


Patterns across the reports

  1. The code has to run, not just exist. Three reports put getting services up and running at the center (01 and 02 for software engineers, 07 for the MLE). Report 04 finished with an end-to-end test, report 06 started by making a provided test pass, and report 08 built live updates one milestone at a time.

  2. Software engineers kept getting the refund workflow. Reports 01, 03, 05 and 06 all name it, from December 2025 to November 2026. The two detailed versions line up closely: in 03 and 06 the starter code already models the workflow as a DAG of nodes, and both moved on to partial refunds and to what happens when a run is slow. The MLE in report 07 expected the refund task and got a different one. The full-stack candidate in report 08 also had a different one.

  3. Candidates picked the AI tool. Report 03 worked in their own AI editor on their own laptop, and report 06 brought a local editor with AI set up. Reports 06 and 08 both say a browser chat alone wasn't enough; 08 names Claude Code, Cursor or Codex as the kind of tool expected. In report 04, the interviewer asked why the candidate chose that tool.

  4. Model choice came up twice. The interviewer in report 04 asked why that model. Report 08 lists model choice among the things being graded.

  5. Speed or care? The reports don't agree. Two candidates felt speed won (02 and 07). One saw fast iteration valued over careful review of AI code, and the other was told to paste the prompt into the AI and move fast. Two others point the other way: report 08 describes a rubric that included checking the AI's output with tests and reviewing PRs, and report 06 says prompting is graded and that leaning too hard on the AI shows. Report 04 couldn't tell what was being measured.

  6. HackerRank carries the files, not the work. In reports 03 and 06 the task or starter code came from HackerRank and the finished code went back there, but the coding itself happened in the candidate's own AI editor.

  7. Plain Code Craft is a separate round. Reports 02 and 03 had a Code Craft phone screen, and report 08 had a CodeCraft onsite round without AI next to the AI one. In all three loops, the AI round came on top of a Code Craft round without AI.

Common questions

What is DoorDash AI CodeCraft?

An onsite hour where you build and run small services live with your own AI tool; software engineers kept getting a refund workflow (01, 03, 05, 06).

Which AI tool should you use for DoorDash CodeCraft?

Candidates used their own AI editors; 08 names Claude Code, Cursor or Codex as the kind expected, and 04 was asked why they chose their tool and model.

Keep practicing

AI coding rounds at other companies

More companies with AI rounds on FastPrep:

Across companies: 2026 AI Coding OAs at 15 Companies: 32 Reported Problems and 2025–2026 AI Coding Phone Screens and Onsites at 43 Companies: 76 Reported Problems.

Sources: Anywhere useful interview info can be found.

0 total reactions. Your reactions: none.

Comments

0

No comments yet — be the first to share what you know.