American Express interview questions with answers (engineering)
Last updated 2026-11-19
American Express hires engineers in India across its technology teams, with data, payments and fraud among the major themes. Candidates commonly describe interviews that mix coding and SQL with practical design questions about data pipelines, APIs and risk, followed by behavioural questions about ownership and decision-making.
The thirteen questions below are practice in that style, with worked answers. They are not past questions. For the online assessment, see American Express coding round questions.
| Typical roles | Graduate engineer or analyst in technology teams (India) |
|---|---|
| Rounds, as commonly reported | Online assessment, technical interviews, managerial or HR round |
| Heavy topics | SQL, DSA, Python or Java, data, APIs, fraud basics |
| Hiring news | See Hiring status below |
Rounds as commonly reported
Candidates commonly report an online coding assessment, then technical interviews and a final conversation with a manager or HR. The company has also run a hackathon, CodeStreet, as noted below. Steps differ by programme and campus.
In technical interviews, expect to code a problem, write SQL on the spot and talk through a design question such as a rate limiter, a pagination scheme or a fraud-scoring flow. Explaining your trade-offs matters as much as the answer.
What to prepare
- SQL including window functions, aggregation and joins on transaction-style tables.
- Python or Java with data structures; generators, collections and complexity.
- Basic data and ML metrics: precision, recall, imbalanced data.
- APIs and system design basics: pagination, rate limiting, idempotency.
- Behavioural stories on ownership and data-driven decisions. See how to crack the HR interview.
Hiring status
For news that changes month to month, see American Express hiring, which is updated from sourced reports. This page keeps only the facts below, each with its source.
Check American Express's own careers site and your campus placement cell for the current cycle.
- CodeStreet 2026 ran online on HackerEarth: registration 13 to 27 July 2026, a prototype round from 7 to 24 August and a finale on 25 August 2026, for teams of one to three. The next edition has not been announced.
- American Express's India technology careers page says applications for its one-year Graduate Full-Time Program open from July to November, with campus hiring run through universities. The page is undated, so treat the window as a pattern, not a 2027 announcement.
American Express interview questions with answers (engineering): questions with worked answers
1 SQL: for each month, find the three customers with the highest card spend.
mediumAggregate spend per customer per month first, then rank within each month using DENSE_RANK or ROW_NUMBER partitioned by month and ordered by total spend descending, and keep ranks up to three. Choose DENSE_RANK to keep ties, ROW_NUMBER for exactly three rows. Truncate dates to month in the inner query. See SQL window functions interview questions.
sqlSELECT month, customer_id, total FROM ( SELECT DATE_TRUNC('month', txn_date) AS month, customer_id, SUM(amount) AS total, DENSE_RANK() OVER (PARTITION BY DATE_TRUNC('month', txn_date) ORDER BY SUM(amount) DESC) AS rnk FROM Transactions GROUP BY DATE_TRUNC('month', txn_date), customer_id ) t WHERE rnk <= 3;2 What is the difference between RANK, DENSE_RANK and ROW_NUMBER?
easyFor salaries 100, 100, 90: ROW_NUMBER gives 1, 2, 3 (always unique, ties broken arbitrarily), RANK gives 1, 1, 3 (ties share a rank and a gap follows), and DENSE_RANK gives 1, 1, 2 (no gap). Use ROW_NUMBER for deduplication, RANK when gaps are meaningful and DENSE_RANK for 'top N distinct values'. Without a deterministic ORDER BY tie-breaker, ROW_NUMBER results can change between runs.
3 In Python, what is a generator and why would you use one to process a huge transactions file?
easyA generator function uses yield to produce values lazily, one at a time, keeping only the current state in memory. Reading a file line by line and yielding parsed records lets you process a file larger than memory in O(1) space instead of loading a list. Generators are single-use and can be chained in pipelines. Mistake: calling list() on the generator, which defeats the point.
pythondef large_txns(path, limit): with open(path) as f: for line in f: acct, amount = line.strip().split(',') if float(amount) > limit: yield acct, float(amount)4 How would you detect suspicious card transactions? Compare rules and machine learning.
mediumRules are transparent and quick to deploy: a spend far above the customer's usual amount, many transactions in a short window, or purchases in two distant locations within minutes. Machine learning models score each transaction using features such as amount relative to history, merchant category, device and time, and catch patterns rules miss, but need labelled data and explanation. In practice both are combined: rules for known patterns and hard limits, a model for a risk score, and human review for the middle band. Mention that fraud data is heavily imbalanced, so accuracy is a misleading metric.
5 A fraud model flags 100 transactions; 80 are truly fraud. There are 200 fraudulent transactions in total. Compute precision, recall and F1.
mediumPrecision is true positives over flagged: 80 / 100 = 0.8. Recall is true positives over all actual fraud: 80 / 200 = 0.4. F1 is the harmonic mean: 2 x 0.8 x 0.4 / (0.8 + 0.4) = 0.64 / 1.2, about 0.533. A high precision with low recall means few false alarms but many missed frauds; the right trade-off depends on the cost of each. Accuracy would look excellent because almost all transactions are legitimate.
6 Design a rate limiter for an API: at most N requests per user per minute.
hardA token bucket stores a token count and last refill time per user. Each request first adds tokens for the elapsed time (capped at the bucket size), then consumes one token if available, otherwise it is rejected. This allows short bursts while enforcing the average rate. Per-user state in a single process is a dictionary; across servers, keep counters in Redis with atomic operations. A fixed-window counter is simpler but allows double the rate across a window boundary.
pythonimport time class TokenBucket: def __init__(self, rate_per_sec, capacity): self.rate, self.cap = rate_per_sec, capacity self.tokens, self.t = capacity, time.monotonic() def allow(self): now = time.monotonic() self.tokens = min(self.cap, self.tokens + (now - self.t) * self.rate) self.t = now if self.tokens >= 1: self.tokens -= 1 return True return False7 How would you design pagination for a transaction-history API that must stay correct while new transactions arrive?
mediumOffset pagination (page=3) skips or repeats rows when new rows are inserted between requests and gets slower on deep pages. Cursor pagination returns an opaque cursor, such as the last seen (timestamp, id), and the next request asks for rows before it, using an indexed ORDER BY on those columns. It stays consistent as data changes and uses an index seek. Include a stable tie-breaker like the id, a sensible maximum page size and a way to detect the end.
8 Explain MapReduce and when you would prefer Spark.
mediumMapReduce splits work into map tasks that emit key-value pairs and reduce tasks that aggregate by key, with a shuffle in between and results written to disk after each stage. Spark keeps intermediate data in memory as resilient distributed datasets or DataFrames and builds an optimised execution plan, which makes iterative and interactive jobs far faster. Choose Spark for iterative analytics, machine learning and SQL-style transformations on large data, and note that both rely on partitioning keys well to avoid skew. Word count is the standard example: map emits (word, 1), reduce sums per word.
9 Sketch the flow of a card payment authorisation and where could it fail?
mediumThe merchant's terminal or website sends the request to an acquirer, which routes it over the card network to the issuer. The issuer checks the card status, available credit and fraud rules and returns approve or decline within a second or two. Settlement happens later in batch. Failures include timeouts (the merchant does not know whether the issuer approved), duplicate submissions, and declines for fraud. Idempotency keys and reversal messages for timeouts are the standard mitigations.
10 When would you denormalise a database, for example in a data warehouse?
mediumNormalisation removes redundancy and update anomalies, which suits transactional systems. Analytical workloads read large volumes and join many tables, so a warehouse often uses a star schema: a central fact table of transactions with dimension tables such as customer, merchant and date, partly denormalised for fast aggregation. The cost is duplicated data and more complex updates, handled by loading in batches. The decision is workload-driven: many writes favour normalised, heavy reads favour denormalised.
11 Why American Express, and why technology?
easyDescribe what attracts you to payments and data: large-scale systems where reliability, fraud detection and customer data come together. Mention a relevant project or course, such as a data pipeline or a model you built. Say what you want to learn in the first year. Keep it grounded in the job description and avoid quoting company statistics you have not read from an official source.
12 Tell me about a time you used data to make a decision.
mediumPick a project where you measured something: for example comparing two algorithms by running time, or using survey results to choose features. State the question, the data you collected, how you checked it was reliable, what the data showed and the decision. Include one limitation of your analysis. Interviewers want to see structured thinking more than a sophisticated tool.
13 Describe a time you took ownership of something outside your assigned work.
mediumChoose a case where you noticed a gap, such as unclear documentation or a flaky test, and fixed it without being asked. Explain why you chose it, how you made sure you were not stepping on others, what you did and what changed. Keep the example proportionate. Ownership in a bank also means escalating the things you cannot fix yourself.
Frequently asked questions
Does American Express ask SQL questions?
Candidates commonly report SQL questions for data-heavy engineering roles, including ranking and aggregation. Practise window functions on transaction-style tables.
What is CodeStreet?
It is an American Express hackathon run online on HackerEarth. The 2026 dates are given in the hiring section with the source.
Are machine learning questions asked?
It depends on the team. For fraud and data roles, expect basic metrics such as precision and recall. Software roles focus on coding and design.
Where can I see current openings?
See American Express hiring for sourced notes and the company's own careers site for live roles.
Sources
Related guides
- American Express coding round questions with solutions
- Visa interview questions with answers (software engineer, India)
- Mastercard interview questions with answers (software engineer)
- PayPal interview questions with answers (software engineer, India)
- SQL Window Functions Interview Questions
- HR Interview Questions: Behavioural Prep for Freshers
More in Company Interview Questions
- Analog Devices Coding Round Questions with Solutions
- Analog Devices Interview Questions and Answers for India Embedded and DSP Roles
- Apple Coding Round Questions with Solutions in Python and Java
- Apple Interview Questions and Answers for SDE Roles
- Arcesium Coding Round Questions With Solutions
- Arcesium Interview Questions and Answers
- Atlassian Coding Round Questions with Solutions in Python and Java
- Atlassian Interview Questions and Answers for SDE Roles
- Barclays coding round questions with solutions
- Barclays interview questions with answers (technology graduates)
- Birlasoft Coding Round Questions with Solutions
- Birlasoft Interview Questions for Freshers