The rubric stays hidden
The interviewer is briefed with weighted criteria it is told never to reveal. You cannot write to the test, because you cannot see the test, exactly like the real thing.
Practise the SRE, platform or engineering manager loop you have booked. One question at a time, follow-ups on what you actually said, and a rubric you never see. Then a second model grades the transcript out of 100 and quotes your answers back to you.
In private beta.
Site Reliability Engineer · senior · Incident Scenario, question 3/8
Q: It's 02:14. Checkout p99 latency has gone from 180ms to 4.2s over the
last eight minutes and the error rate is at 6%. Payments and search look
normal. You're on call and you've just opened the laptop, what do you do
first, and what are you looking at?
│ First I'd declare a Sev-2 and page the checkout owner, then check
│ whether this correlates with a deploy. I'd look at the deploy
│ timeline before touching anything. Impact first: what fraction of
│ checkouts are failing, and is it all regions or one?
│
│
ctrl+d submit · ctrl+c quit
Practising in a general-purpose chat has two failure modes: it is relentlessly encouraging, and it forgets what it was assessing. mockterview is built the way a real loop is, brief, rubric, panel, debrief, and keeps the assessment hidden until the end.
The interviewer is briefed with weighted criteria it is told never to reveal. You cannot write to the test, because you cannot see the test, exactly like the real thing.
A higher-quality model runs the conversation at temperature 0.7. A separate, cheaper grader scores the finished transcript at 0.2. The grader is never the model you were talking to.
Every criterion comes back with a score out of ten, its weight, and a comment that quotes your own words, so the feedback can be argued with.
No account to make, and nothing that stops working when the Wi-Fi does. Beta members get the binary, every bundled role and rubric, and the updates as they land. Every feature, in detail.
Asks one question at a time, follows up on vague answers, and escalates depth for mid, senior, and staff candidates.
Role-plays the production system: symptoms only appear when you go looking, and a complication lands partway through.
Per-criterion scores out of ten, normalised into one weighted result out of 100, with written feedback per criterion.
Every session is stored locally. history shows your average per interview
type and the delta from your first run to your latest.
Drop in a new role or interview format as a YAML file. No code, no rebuild, no waiting on a vendor roadmap.
Talks to any OpenAI-compatible endpoint. Point it at a local runtime and the whole thing works offline.
No CGO, no runtime, no daemon, no container required. Copy one file onto the machine and run it.
Transcripts and scores live in one file you own. Back it up, inspect it with any SQLite client, or delete it in one command.
Nothing phones home. The only secrets involved are your own endpoint URL and key,
kept in a local .env.
Pick a role, a seniority, and an interview format, or pass them as flags and skip straight in.
Up to eight questions, multi-line answers, Ctrl+D to submit. The interviewer follows the thread you actually opened.
The grader scores the transcript against the hidden rubric and writes feedback that cites what you said.
Re-run the same format next week and watch the per-type average move.
Every panel below is verbatim output from the tool, the picker, a live incident interview, the graded result, and the history commands.
Select a role
> Engineering Manager
Platform Engineer
Site Reliability Engineer
↑/↓ move · enter select · ctrl+c quit
Role, then seniority, then format, or pass --role,
--seniority and --type and go straight to question one.
Site Reliability Engineer · senior · Incident Scenario, question 3/8
Q: It's 02:14. Checkout p99 latency has gone from 180ms to 4.2s over the last eight minutes and
the error rate is at 6%. Payments and search look normal. You're on call and you've just
opened the laptop, what do you do first, and what are you looking at?
│ First I'd declare a Sev-2 and page the checkout owner, then check whether this
│ correlates with a deploy. I'd look at the deploy timeline before touching
│ anything. Impact first: what fraction of checkouts are failing, and is it all
│ regions or one?
│
│
│
ctrl+d submit · ctrl+c quit
The scenario is generated for your role and seniority. Symptoms are revealed only in response to what you ask for.
Interview complete
Overall: 72.5 / 100
Rubric
Triage & prioritisation 8.0/10 (weight 20)
Declared severity and paged the service owner before debugging, "Impact first: what
fraction of checkouts are failing" is the right opening move.
Diagnosis & hypothesis testing 7.5/10 (weight 30)
Correlated with the deploy timeline, but you went to the database before ruling out
the connection pool you yourself flagged.
Mitigation & recovery 7.0/10 (weight 20)
Rollback was the safe call. You did not mention blast radius of draining the region.
Communication & coordination 7.0/10 (weight 20)
Good stakeholder update cadence; you stayed hands-on-keyboard as IC instead of delegating.
Follow-up & prevention 6.0/10 (weight 10)
Postmortem mentioned, but no concrete action items or guardrail proposed.
Feedback
You lead with impact and you stop the bleeding; that reads as senior. The gap is discipline
under your own hypotheses: you said "it smells like connection pool exhaustion" at Q3 and then
never tested it, which cost you four minutes of the simulated incident. Next run, write the
hypothesis down and name the signal that would falsify it before you move on.
session a1b2c3d4 · saved · press enter to exit
Weights come from the rubric file, so a staff-level system design and a behavioural screen are scored on different things.
$ mockterview history
ID DATE ROLE SENIORITY TYPE SCORE
----------------------------------------------------------------------------------------
a1b2c3d4 2026-09-09 21:12 sre senior incident-sc… 72.5
7c1de9f0 2026-09-08 19:12 platform-engineer staff system-desi… 78.0
3b0a55e1 2026-09-07 17:12 sre senior incident-sc… 64.0
e4f2a7b9 2026-09-05 19:12 engineering-manager senior behavioral 69.5
9f8e7d6c 2026-09-02 19:12 sre senior incident-sc… 58.0
Progress by type (graded sessions):
system-design avg 78.0 over 1
incident-scenario avg 64.8 over 3 (first 58.0 → latest 72.5, +14.5)
behavioral avg 69.5 over 1
The trend line is per interview type, so improving at incidents is not hidden by a good behavioural run.
$ mockterview show a1b2c3d4
Session a1b2c3d4
Role: sre Seniority: senior Type: incident-scenario
Started: Wed, 09 Sep 2026 21:12:00 UTC Status: graded
Interviewer model: auto → qwen3-coder:30b Grader model: auto → llama3.1:8b
Transcript
------------------------------------------------------------
Q1: It's 02:14. Checkout p99 latency has gone from 180ms to 4.2s over eight minutes,
error rate 6%. What do you do first?
A1: Declare a Sev-2, page the checkout owner, and check the deploy timeline before
touching anything.
Q2: There was a deploy 20 minutes ago, a config change to the connection pool.
Rolling back takes 6 minutes. What now?
A2: Roll back immediately; it's the cheapest reversible action and the correlation is
strong. I keep investigating while it drains.
Score: 72.5 / 100 (graded by llama3.1:8b)
------------------------------------------------------------
Triage & prioritisation 8.0/10 (weight 20)
Severity and ownership established before debugging.
Diagnosis & hypothesis testing 7.5/10 (weight 30)
Strong deploy correlation; the pool hypothesis was never tested.
Feedback
You lead with impact and stop the bleeding. Write your hypothesis down and name what
would falsify it.
Full transcripts are replayable, so a coach or a hiring manager can read the session back and disagree with the grader.
Roles and rubrics are plain YAML. Copy one, change the weights to match your own
scorecard, and every interview you sit is graded against it. The file's
id is what you pass to --role and --type.
id: system-design
name: "System Design"
# Short brief the interviewer uses to open and steer.
# It never reveals the criteria or scores below.
interviewer_brief: >
Open with a deliberately ambiguous, realistic design
prompt appropriate to the candidate's role and
seniority. Ask ONE question at a time. Follow up on
vague answers by asking for specifics. Do not
lecture; do not reveal how you are assessing them.
# Hidden rubric. Consumed by the grader only.
criteria:
- id: requirements
name: "Requirements & scoping"
weight: 15
- id: architecture
name: "High-level architecture"
weight: 25
- id: tradeoffs
name: "Trade-offs & bottlenecks"
weight: 25
Run the same format three times in a week before an SRE or platform interview and watch the score move. It does not go easy on you because you sound confident, and it will not tell you an answer was great when it was a six.
Staff expectations are written into the role files, more scope, more ambiguity, second-order effects. Sit the staff-band version of your own discipline and read the rubric comments on where you still sounded senior.
Coming back after a break, or moving from dev into SRE, the first two interviews are where you learn what you forgot. Spend those here, at 3pm on a Tuesday, instead of on the company you actually wanted.
Every learner gets their own history file showing progress across the course, and you get rubrics that match what you actually teach. Nobody counts the sessions, in week one or week six.
Encode your real scorecard as a rubric file and have each interviewer take the loop they are about to give. Disagreements surface against something written down rather than in the debrief.
Practice means being bad at it first. Point it at a model on your own machine and no account, no vendor and no future employer ever sees the session, the transcript is a local file you can delete in one command.
mockterview is in private beta. Request access and you will hear when there is a seat.
No. Beta members get it as a binary. The role and rubric files ship as editable YAML, so you can change how it interviews without access to the source.
One machine and an OpenAI-compatible model endpoint. That can be a local runtime on
your laptop, a model server inside your network, or a hosted provider behind a proxy,
the endpoint URL and key live in a local .env file.
Transcripts are written to a local SQLite file and are sent only to the model endpoint you configure. Point that at a model inside your own perimeter and nothing leaves it. mockterview itself has no telemetry and no accounts; we cannot see your sessions, in the beta or after it.
Yes; that is the point of the YAML rubrics. Criteria, weights and the interviewer brief are all editable, and a new interview format is a new file rather than a code change. That applies whether "our" means your company or just the kind of loop you keep failing.
No, and the licence does not permit it. mockterview is built for a person to practise on themselves, or for an interviewer to sit a loop before they run it. It is not built, tested or offered as a tool for screening, ranking or selecting job applicants, and its scores must not inform a hiring decision about anyone other than the person who sat the interview. Using it that way would put you in scope of the EU AI Act's rules for high-risk employment systems, as the deployer.
No. Both the interviewer and the grader are AI systems: two separate language models, as described above. Every question, score and comment you receive is generated, not written by a human reviewer.
Your application goes on the beta list. When your place opens you get the binary, the install steps and every bundled role and rubric by email.
Tell us the loop you are preparing for. Beta members get the binary, the install steps and every bundled role and rubric.