Singapore version. Adds what Singapore ML interviews probe: using personal data to train models under PDPC's AI guidelines, data minimisation, and human oversight from the Model AI Governance Framework. US version

Machine Learning Engineer Interview Prep (Singapore)

Machine learning interview prep for Singapore: training data under the PDPA business improvement exception, AI governance from IMDA and PDPC, and human oversight.

The interactive tool loads here when JavaScript runs.

About this tool

What is ML Engineer Interview Prep?

The algorithms, metrics and system design on this page are the same anywhere. What Singapore interviewers at banks, platforms and government-linked companies add is the data question: can you use this personal data to train the model at all? PDPC's Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems, issued on 1 March 2024, answer it in PDPA terms.

Organisations can use personal data with meaningful consent, or rely on exceptions. The business improvement exception covers using data already collected under the PDPA to improve existing products or develop new ones, understand user behaviour and personalise services; the guidelines name recommendation engines and HR systems that shortlist candidates as examples, and say it can cover testing and bias assessment. The research exception covers commercial research without a product roadmap. Either way the guidelines expect data minimisation, and policies on when to train on anonymised or pseudonymised data.

The Model AI Governance Framework (second edition) from IMDA and PDPC frames the other common question: how much human involvement a decision needs. It describes human-in-the-loop, human-over-the-loop and human-out-of-the-loop, chosen by the severity and probability of harm. Fraud and credit models are the usual interview examples.

  • Four frameworks: ML system design, model evaluation, data pipelines and MLOps
  • 32 core concepts across theory, data, evaluation and production, with progress tracking
  • Four ML system design problems with problem-specific considerations
  • Editable, copyable solution outline
  • 30 sample questions in five groups
Read the guide: Engineering Manager, ML Engineer and Designer Interviews Explained

Updated

How to use ML Engineer Interview Prep

  1. Read each framework on the ML Frameworks tab with View Framework.
  2. On Core Concepts, tick each topic once you can explain it and give an example; the header percentage tracks your coverage.
  3. On the System Design tab, choose a problem in Select ML Problem Type and describe the context in Problem Description, for example "fraud scoring for card payments at 2,000 transactions per second".
  4. Click Generate ML Solution, then rewrite the outline in the text box as your own design and use Copy to keep it.
  5. Practise answers from the Sample Questions tab out loud, especially the system design group.

Worked examples

Recommendation system for an existing app, answered for Singapore

Input

Case Study: Recommendation System. Interviewer asks: "Can we train on users' browsing history?"

Output

Answer outline: the data was collected under the PDPA for this service; the business improvement exception can cover enhancing the existing product; use only the attributes needed; pseudonymise for training; tell users how recommendations use their data

PDPC's guidelines give "recommendation engines in social media services that offer users content more aligned to their browsing history" as an example of the exception. Naming data minimisation and notification shows you know the exception is not a blank cheque.

Fraud model and the level of human oversight

Input

Case Study: Fraud Detection. Interviewer asks who reviews blocked payments

Output

Answer outline: automated scoring at payment speed (human-over-the-loop, with people monitoring and able to adjust thresholds); a human-in-the-loop review queue for account closures or large blocked transfers

The Model AI Governance Framework ties the level of human involvement to the severity and probability of harm, and notes that a human in the loop is impractical where decisions must be made at high speed. Splitting the decision by impact is the answer interviewers look for.

Background and insight

ML engineer loops typically combine a standard coding round, an ML fundamentals round, an ML system design round and behavioral questions. The system design round is where most candidates lose points. Interviewers want to hear the business objective turned into an ML objective, a simple baseline before anything clever, where labels come from, how offline metrics relate to the online metric you actually care about, and what happens when the data drifts.

Metric choice is a common trap. On a fraud dataset where 0.1% of transactions are fraudulent, a model that predicts "not fraud" every time has 99.9% accuracy and is useless. Precision (of the flagged cases, how many were fraud), recall (of all fraud, how much was caught) and the precision-recall curve tell you far more. For ranking, use metrics such as NDCG or MRR that reward putting relevant items near the top.

Leakage and skew cause most real-world failures. Leakage means a feature carries information that will not exist at prediction time, such as a "refund issued" flag in a fraud model, and it produces offline scores that collapse in production. Training-serving skew means features are computed differently in the training pipeline and the live service. Mentioning both, and how a feature store or shared feature code prevents skew, signals production experience.

Common use cases

  • Preparing for an ML system design round

    Practise all four problems with the outline until you can cover every section in 40 minutes.

  • Moving from data science to ML engineering

    Focus on the Production ML Systems concepts and the MLOps framework, which data science interviews rarely cover.

  • Refreshing fundamentals

    Use the theory questions to check you can still explain regularization, bias and variance, and backpropagation clearly.

  • Running interviews

    Use the practice problems and their considerations as a shared rubric for an ML design round.

Tips and limitations

  • When asked about training data, say which PDPA basis you rely on: consent, the business improvement exception or the research exception.
  • Mention bias assessment as a legitimate use of personal data; PDPC's guidelines say the business improvement exception could apply to it.
  • Describe the human oversight model by risk: in, over or out of the loop, as the Model AI Governance Framework names them.
  • Keep data minimisation in your feature engineering answer: only the attributes the model needs.

Sources and further reading

  1. Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems, Personal Data Protection Commission. Issued 1 March 2024: business improvement and research exceptions, data minimisation, notification
  2. Model Artificial Intelligence Governance Framework, second edition, IMDA and PDPC. Human-in-the-loop, over-the-loop and out-of-the-loop
  3. Personal Data Protection Act 2012, section 26D, Singapore Statutes Online. 3 calendar days to notify the PDPC
  4. Guide on Managing and Notifying Data Breaches Under the PDPA, Personal Data Protection Commission. Notify the Commission as soon as practicable and no later than three calendar days after determining a breach is notifiable

Frequently asked questions

Can I train a model on customer data in Singapore without fresh consent?

Sometimes. PDPC's 2024 advisory guidelines say the business improvement exception lets organisations use personal data they collected in accordance with the PDPA to improve or develop products, understand behaviour and personalise services, which can include training, testing and bias assessment of AI systems. Data minimisation and accountability policies still apply.

What is the difference between the business improvement and research exceptions?

Per PDPC, the business improvement exception fits when you are enhancing an existing product or developing new ones in your business; the research exception fits commercial research to advance science and engineering without a product development roadmap.

What does human-in-the-loop mean in the Model AI Governance Framework?

It is one of three levels: a human makes the final decision (in the loop), a human supervises and can intervene (over the loop), or the system decides without human involvement (out of the loop). The framework chooses among them by the severity and probability of harm.

Do I need to tell users that an AI system uses their data?

Where you rely on consent, yes: PDPC says the notification obligation means telling users the purpose and the types of data, and suggests notification pop-ups or accessible written policies. Organisations are encouraged to explain the product function that needs the data.

Which PDPA rule applies if training data leaks?

The breach notification duty: if the breach is assessed as notifiable, the PDPC must be told within 3 calendar days of that assessment. Pseudonymised training sets and access logs reduce both the harm and the work of that assessment.

Is there an in-depth guide to ML Engineer Interview Prep?

Yes. “Engineering Manager, ML Engineer and Designer Interviews Explained” covers the background, how it works, common pitfalls and worked examples. Read the full guide.