An AI interviewer for SQL, coding, product and case rounds. It follows your query, code and notes, asks why, and ends with a scorecard that quotes you. Free, no sign-up.
What is the unit you're counting at each step, and how will you enforce step order?
Sessions, not events. Each step uses the first time the previous one happened…
3 of 4 datasets match. What's special about the zero-denominator one?
WITH eligible AS (
SELECT s.session_id, s.platform
FROM sessions s
WHERE s.platform IN ('android', 'ios')
), menu AS ( … )
SELECT platform,
ROUND(100.0 * orders
/ NULLIF(menu, 0), 1)This is not a question bank. It isn't a leaderboard, a streak counter, or a list of questions someone swears they were asked at Google.
It's a room, a question, and an interviewer who is actually listening.
That's the whole idea. The rest of this page explains why.
Interviews score how you reach an answer out loud, with someone in the loop. Silent practice doesn't train that.
Most preparation happens in silence: a problem, an editor, a green checkmark. Real interviews don't work that way.
Someone is watching you think. They interrupt. They ask why you chose a hash map, what happens when the list is empty, which number you would check first.
The skill being tested isn't only reaching the answer. It's reaching it out loud, under gentle pressure, with another person in the loop.
Practice without a listener, rumours dressed as facts, and scores without evidence. Margin answers each one.
Grinding alone builds pattern recall. It doesn't build the habit of narrating a decision while someone asks “why?”
In Margin a voice interviewer asks one question at a time and follows what's in your workspace: the failing test, the query grain, the branch of your issue tree.
“Asked at Amazon” tags, frequency charts with no source, casebooks passed around without dates. They feel precise and are rarely checkable.
In Margin every question carries one of four labels, and a label is only used when the evidence exists.
No “asked at” tags or frequency statistics are invented. The current catalog has no reported or official questions, because none is backed by a dated source yet. Company context comes from public sources with review dates, and fictional numbers are labelled fictional.
“7 out of 10” tells you how it went, not why. Without the why, you practise the same mistakes.
In Margin each interview type has its own rubric, and every score quotes what you said, wrote, or ran.
None of this needs to be complicated. It needs to be honest.
Pick a question and a tone. Work in a real workspace while the interviewer follows along. Leave with evidence, not a vibe.
Start from SDE, ML & Data, Product, or Consulting, or from the company you're targeting. Every brief shows the skills tested, the workspace you'll use, and where the question comes from, with the date its source was reviewed.
Supportive, neutral, or challenging: tone changes the wording and the pacing. It never changes the tests, the evidence standard, or the scorecard.
Python with hidden tests. Real PostgreSQL with edge-case datasets. Notes for cases. A canvas whose text stays editable.
The interviewer sees the latest version of your work, and only speaks when you run, check, or ask.
SQL is scored on query correctness and relational reasoning, system design on architecture and trade-offs, consulting on structure and arithmetic. Objective results are pinned; everything else cites evidence.
Company-inspired practice Original brief using an e-commerce flash-sale context.
Flipkart · fictional numbers · reviewed 2026-09-27Same for every tone: voice · tests · evidence standard · scorecard
Interviewer Walk me through your approach before you run it.
Interviewer Why DENSE_RANK here, and not ROW_NUMBER?
Interviewer What would make you reverse that recommendation?
Interviewer Stale after 20 seconds. What does the customer see then?
def two_sum(nums, target):
seen = {}
for i, v in enumerate(nums):
if target - v in seen:
return [seen[target - v], i]
seen[v] = iSELECT d.name AS department, r.name, r.salary
FROM (
SELECT *, DENSE_RANK() OVER (PARTITION BY dept_id
ORDER BY salary DESC) AS band
FROM employee) r JOIN department d ON d.id = r.dept_id
WHERE r.band <= 3| department | name | salary |
|---|---|---|
| Data | Asha | 182000 |
| Data | Ravi | 175000 |
| Design | Meera | 140000 |
## Objective
Break-even within 24 months?
## Analysis
420k × 18% × 3.2 = 241,920 orders/month · ₹9.6 contribution/order
## Recommendation
Decline full launch; two-store pilot with kill criteria.
Scroll through the stages, or pick one. Each plays out live, cursor and all.
From quick drills to full mock rounds. Start from a role, or from the company you're targeting.
DSA in Python with hidden tests, and system design on a canvas.
10 DSA · 10 system design20 questions 02ML & Data AnalystPostgreSQL exercises checked against edge cases, and applied ML cases.
5 SQL · 5 ML cases10 questions 03Product AnalystMetric trees, experiments, funnels, and cohort SQL.
8 cases · 2 SQL10 questions 04Consulting & GuesstimatesCandidate-led cases with exhibits, and estimates with no single right number.
5 cases · 5 estimates10 questions 05AI EngineerSpoken rounds on AI product design, LLM systems, projects, behavioural questions, and full mocks.
Product · Projects · HR · Mocks41 questionsPick one question. Say your first thought out loud.
50 questions, every one with a source. Filters live in the address bar, so any view can be shared.
Loading questions…
Upload one resume, add an optional role or job description, and Margin will prepare one grounded interview. Margin keeps your resume's text for up to 90 days so you can practise again; the uploaded file itself isn't kept.
Each collection pairs public context with practice questions. Labels say exactly what is known about each question; Margin never invents "asked at" tags or frequency statistics.
Your reports are saved with Margin and linked to this browser, so they show up only here. Open one to see the verdict, the feedback, and what to practise next. How we handle your data.
Tone changes wording and pacing only, not the voice, evidence standard, tests, or scorecard. Hints are given only when you ask, and are recorded.
Run your code when you’re ready. The interviewer will see the result.
Run query executes against the visible sample. Check answer also runs the additional datasets.
Check answer compares results (not query text) on every dataset.
Notes sync after a short pause without interrupting the interviewer. “Discuss my work” asks for feedback.
Pen strokes and text boxes stay editable objects. Click with Text to add a box; use Select to move, resize, or edit. The interviewer’s notes are labelled and never change your work.