Software engineering
AI interviews for software engineers
Talk through your real engineering work with an AI interviewer, then see your results first and choose who sees them. Below: what strong answers show, the 5 practice questions, and a worked example.
- Free for experts
- About 15 minutes, whenever you like
- Type or speak. No camera needed
Retake it up to 3 times: you keep every attempt, and choose which one represents you.
Who it's for
The work it covers
Building and running software: web, mobile, backend, infrastructure, security and AI applications.
Example roles
- Backend engineer
- Frontend engineer
- Mobile engineer
- Site reliability engineer
- Security engineer
- Full-stack AI engineer
- ML platform engineer
- AI product engineer
- Data / infra engineer
Name your exact specialty when you apply, and the questions follow it.
Topics the interviewer draws on in software engineering
- A production system they owned and what made it hard
- Debugging a failure or outage from first alert to fix
- A design trade-off between speed, reliability, cost and maintainability
- Scaling a system or codebase as usage grew
- Keeping quality high with tests and code review under a deadline
- Explaining a technical risk to people who are not engineers
- Working with product and design on what to build
- A technical decision they would make differently now and what they learned
Scoring
What a strong answer shows
Software engineering is scored on the evidence rules below, area by area. The strongest answers name one system, what made it hard, the options weighed and a measured result.
The four scored areas
Problem solving
Approach, debugging, edge cases
Counts as evidence when the candidate describes a specific problem, its cause, a fix, a trade-off, or a measured result.
Architecture
System design, components, data flow
Counts as evidence when the candidate describes how a system is built: components, data flow, storage, queues, APIs, scaling or consistency choices.
Communication
Clarity, structure, ability to explain
Counts as evidence when the candidate describes explaining, documenting or presenting work to other people (for example a design doc, a runbook, or walking a team through a change). Answers that respond to the question asked, in a clear order, with a concrete example, also count, but only up to 80 on their own. Grammar, word choice and fluency are never judged.
Collaboration
Teams, reviews, feedback
Counts as evidence when the candidate describes working with other people: teammates, stakeholders, reviews, mentoring, feedback or cross-team work. Saying "we" is not, by itself, collaboration evidence.
An area is scored only when the AI can quote you; otherwise it shows Not enough evidence, not a low score. How AI interviews are scored
Practice
The 5 practice questions
Questions like these come up in the interview. The free practice room asks a different set, written for your role and level. The real interview follows up on your own answers instead.
- Question 1
Tell me about a production system you owned, what made it difficult, and what you personally changed.
- Question 2
What trade-off did you make between speed, reliability, cost, or developer experience, and how did you know it was the right trade-off?
- Question 3
Imagine that system starts failing under a 10x traffic spike. Walk me through the first three signals you would inspect and the mitigation you would ship first.
- Question 4
How would you explain an incident and your recovery plan to a non-technical manager without hiding the technical risk?
- Question 5
What would you improve if you had one more week, and what evidence would prove the improvement worked?
No account. Each answer gets a quick rule check in your browser.
Worked example
One answer, read by the rules
Made up by Aethel to show how the published rules read one answer: not a real candidate, and not scored by the AI. A real score can differ.
The practice question
Tell me about a production system you owned, what made it difficult, and what you personally changed.
The made-up answer
I owned the notification service at a food delivery startup, about 40 engineers, sending roughly 3 million push and SMS messages a day. What made it hard was that every other service called it synchronously over HTTP, so when our SMS provider slowed down, checkout requests timed out with it. I changed it so callers write the message to a queue and get an acknowledgement in a few milliseconds, and a separate pool of workers does the sending, with retries and a dead-letter queue. I looked at three options. Just raising the timeouts would have hidden the problem and still tied checkout to a vendor's latency. Moving everything to a full event bus would have fixed it, but it was a quarter of work and a new system for the on-call rotation to learn. A queue table in our main database was tempting, but at peak write rates it would have competed with orders for locks. So I used the managed queue we already ran for billing, which on-call already knew how to operate. The next time the SMS provider slowed down, checkout p99 latency stayed at about 350 milliseconds, where before it had hit our 9 second timeout.
Marked words are the quotes the rules would look at.
How the rules read it
The situationLevel 2 of 2
A specific case with where it was and how big: one service, a company of about 40 engineers, about 3 million messages a day.
The reasoningLevel 3 of 3
Two or more options compared, with why each was rejected: longer timeouts and a full event bus each lose for a stated reason, and a third, a queue table in the main database, is ruled out next.
The resultLevel 2 of 3
Measured: p99 latency during a real provider slowdown, before and after.
Together the three ratings give 7 of 8 points, which is 91 on the fixed scale. Its longest complete quote has 45 words, so the quote rule does not lower it. It is evidence mainly for Architecture.
What would raise it
Following the result over time, for example checkout timeouts over the next two quarters, or another team adopting the queue pattern, would raise the result to 3 and the area to 8 points.
The same story, told vaguely
I worked on a notification service that other services depended on. It had reliability problems whenever a vendor was slow, so I moved it to a queue with background workers and retries. I considered a few approaches and picked the one that made the most sense for the team. Afterwards it was much more reliable and checkout stopped having problems.
A case without its context: no company, no scale and no message volume, so situation is 1 and the area is at most 60. The options are not named, and "much more reliable" is a vague result.
A whole result: the sample software and data result, with a made-up candidate. More examples: strong interview answers.
Company interview links
Roles companies can set up
Starter templates a company can use for an interview link, and what each one steers the interviewer toward.
Engineering manager
Focus: Leading and growing an engineering team, planning and delivery, technical decisions, hiring and feedback
Frontend engineer
Focus: Building web interfaces, managing state, page speed, accessibility
Mobile engineer
Focus: Building iOS or Android apps, app performance, offline support, releasing through the app stores
QA automation engineer
Focus: Test strategy, automated tests, fixing flaky tests, release quality, working with developers
Security engineer
Focus: Threat modeling, secure code review, finding and fixing vulnerabilities, cloud and access security, incident response
Senior backend engineer
Focus: API design, databases and data modeling, running services in production, code review
Site reliability engineer
Focus: Running services in production, monitoring and alerting, incident response, infrastructure as code, deploy pipelines
FAQ
Questions people ask
Should I list the languages, frameworks and tools I use?
Only when they explain a decision. The areas are scored on what your answers show: a problem with its cause and fix, how a system is built, explaining work to people, and working with others. Saying you chose a queue over longer timeouts, and why, is evidence. A list of tools with no decision behind it is not.
Can I talk about an outage I caused?
Yes. A specific failure is often your strongest problem solving material: what broke, how you found the cause, what you shipped, and what changed afterwards. The scoring looks for a specific case with where it was and how big, your reasons, and a measured result. It does not reward a spotless record. Leave out anything you are not allowed to share.
How much detail should I give about scale?
Enough to place the case: how many users, requests, services or engineers, and roughly how much traffic or data. Without a specific case named with where it was and how big, an area scores at most 60. Rounded numbers are fine, and you can describe the system in general terms.
Is there live coding or a LeetCode-style problem?
No. Nobody writes or runs code. The interview is a spoken or typed conversation about systems you built, broke and fixed, closer to the system design and behavioral rounds than to a coding round. Add your specialty, such as frontend, mobile, infrastructure or security, when you apply, and the questions follow it.
Do I have to share confidential details?
No. Talk about systems in general terms, and leave out secrets, credentials and code you are not allowed to share. The interview is about your decisions and reasoning.
How to prepare for an AI interview
Guides for a company's AI interview:Mercor AI interviewmicro1 AI interviewHireVue interview
What this interview cannot do
It is one conversation of about 15 minutes, so it covers only part of what you can do, and we have not tested whether results predict job performance. The AI can misread an answer, which is why every score shows its quotes and people make the decisions. It is not a certification or license check. No independent bias audit has been done. Software engineering was one of the 8 fields in our own test (not an audit) of whether non-native English grammar changes scores.
Your interview, when you are ready
Hiring software engineers?
Interview links for software engineering, and how to check a result a candidate sends you.