New opportunity

Al Engineer, Large Language Model

Shopee · Singapore

About this role

Shopee seeks an engineer in Singapore to research and develop its Compass ultra-large-scale multilingual foundation model, with a focus on pre-training, alignment and trustworthy large language models.

Responsibilities

Research and implement multilingual pre-training and alignment methods, including Mixture-of-Experts training, Instruction Pretraining, supervised fine-tuning (SFT) and RLHF.
Improve the explainability and safety of large language models. Compare emerging technologies for potential use in business scenarios, and run experiments to assess model performance and identify improvements.
Collaborate with the team using statistical, scripting and programming expertise.

Requirements

Doctorate in Computer Science, Information Technology, Programming & Systems Analysis, or a related discipline.
Strong coding, data-structure and basic algorithm skills, including proficiency in Python and PyTorch.
At least 1 year of research experience with the principles and training methods of leading LLMs, such as GPT or LLaMA, plus research experience in text generation or dialogue systems.
Strong problem-solving skills for large-model training and applications, and the communication and collaboration skills to explore new technologies with the team.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

PythonPyTorchLarge Language ModelsMultilingual pre-trainingMixture-of-ExpertsInstruction PretrainingSupervised fine-tuningRLHFModel alignmentAI safetyModel evaluationStatisticsScriptingData structuresAlgorithmsText generationDialogue systemsProblem solvingCommunicationCollaboration

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Multilingual foundation-model training
The role contributes to ultra-large-scale multilingual pre-training and alignment for Shopee’s Compass model. Prepare to explain research choices for multilingual data, training objectives and model behavior, distinguishing work you have done from approaches you would propose.
Implementing and analyzing training methods
The posting names Mixture-of-Experts training, Instruction Pretraining, SFT and RLHF. Prepare to discuss how you would implement and compare relevant methods in Python and PyTorch, including how you would diagnose a training problem.
Trustworthy model behavior
Improving LLM explainability and safety is an explicit duty. Prepare to describe how you would evaluate these properties alongside text-generation or dialogue quality, rather than treating them as separate claims.
Evidence-led technology decisions
The role calls for comparing frontier methods for business scenarios and experimenting across AI models to find improvements. Prepare to show how you would choose baselines, interpret results and communicate a recommendation to collaborators.

Likely challenges to think through

Scaling without losing experimental clarity
A likely difficulty in ultra-large-scale training is separating an algorithmic improvement from effects of data, compute or implementation choices. Prepare a plan for controlled comparisons and for investigating training or inference efficiency bottlenecks; the team describes these as research aims, not confirmed problems in this role.
Balancing multilingual gains and uneven performance
Multilingual and cross-cultural objectives may make an overall improvement obscure weaker behavior in particular languages or contexts. Prepare to explain how you would inspect results across language groups and decide whether a trade-off is acceptable.
Reconciling capability with safety
Pre-training and alignment work may improve generation quality while changing safety or explainability. Prepare to describe evaluations that reveal such trade-offs and how you would present inconclusive findings to the team, rather than assuming one method improves every dimension.

Practice your answers

Technical practice

How would you design multilingual pre-training for a foundation model?

What to coverThe role includes ultra-large-scale multilingual pre-training, and the team describes multilingual data from emerging markets, particularly Southeast Asia. Preparation suggestion: Discuss data quality, language balance, cross-cultural coverage and how you would evaluate performance across languages.

How would you use scaling laws to plan a model-training experiment?

What to coverThe team aims to develop scaling laws to guide model planning. Preparation suggestion: Explain what you would measure across model, data and compute scales, and how uncertainty in early results would affect the next experiment.

What trade-offs would you examine when training a Mixture-of-Experts model?

What to coverMixture-of-Experts training is named in the job description. Preparation suggestion: Address expert routing, load balance, compute efficiency and the possibility that gains vary by language or task.

How do Instruction Pretraining and supervised fine-tuning differ?

What to coverThe role explicitly includes both Instruction Pretraining and SFT. Preparation suggestion: Explain their objectives, data requirements and how you would test whether each improves instruction-following without weakening other capabilities.

How would you evaluate an RLHF alignment approach?

What to coverRLHF is listed among the role’s alignment methods. Preparation suggestion: Cover preference-data quality, reward-model limitations, changes in helpfulness or safety, and checks for unintended behaviour.

How would you investigate an unsafe or hard-to-explain model response?

What to coverThe job calls for improvements to LLM explanation and safety. Preparation suggestion: Describe how you would reproduce the response, inspect relevant evidence, distinguish a plausible explanation from a verified one and test a proposed fix.

How would you compare models for text generation or dialogue?

What to coverThe role requires research experience in text generation or dialogue systems and involves experiments comparing AI models. Preparation suggestion: Discuss task-specific metrics, human assessment, multilingual test coverage and how you would account for variability.

How would you diagnose a large-model training bottleneck?

What to coverThe requirements emphasize solving large-model training problems, and the team identifies training and inference efficiency as research goals. Preparation suggestion: Explain how you would isolate a bottleneck with measurements before weighing changes against model quality and implementation cost.

What would a reliable PyTorch experiment implementation include?

What to coverThe posting calls for Python and PyTorch proficiency, coding ability, data structures and basic algorithms. Preparation suggestion: Discuss correctness checks, reproducible runs, efficient data handling and how you would detect an implementation error that skews model comparisons.

How would you assess an emerging LLM technique for a business scenario?

What to coverThe job asks for comparisons of frontier technologies for potential business use. Preparation suggestion: Define a relevant baseline and success criteria, then weigh capability, safety and computational cost without assuming a particular Shopee use case.

Behavioural practice

Tell me about a time you solved a difficult model-training or application problem.

What to coverThe requirements emphasize deep problem analysis and solving in large-model training and applications. Preparation suggestion: Use a real situation, explain your diagnosis and actions, and report the result and remaining limitations.

Describe a research experiment that contradicted your expectations.

What to coverThe role involves testing model performance and identifying improvements. Preparation suggestion: Give a real situation, explain how you checked the result and changed your approach, and state what the team learned.

Tell me about a time you explored a new technique with collaborators.

What to coverThe posting calls for teamwork and joint exploration of new technologies. Preparation suggestion: Describe the situation, your contribution to the comparison or implementation, and the outcome of the team’s decision.

Describe a time you communicated a technical trade-off to others.

What to coverThe role requires communication and comparison of technologies for potential business scenarios. Preparation suggestion: Use a real example to show how you presented evidence and uncertainty, handled differing views and reached an outcome.

Sharpen your knowledge

Topics to prepare

Multilingual pre-training and scaling
Posting: Compass is an ultra-large-scale multilingual foundation model; the team describes work on data strategies, scaling laws, and training efficiency. Revision exercise: sketch how you would compare two multilingual data mixes while holding training compute constant, and explain what results would change your next experiment. This is a preparation exercise, not a stated hiring task.
Mixture-of-Experts training
Posting: Mixture-of-Experts model training is named in the role. Revision exercise: explain expert routing, load imbalance, and the trade-off between active computation and total parameters; then propose a controlled comparison with a non-MoE model.
Instruction Pretraining and alignment
Posting: Instruction Pretraining, supervised fine-tuning (SFT), and RLHF are explicitly named. Revision exercise: map the purpose, training signal, and likely failure modes of each stage, then explain how you would test whether an alignment change improved responses without obscuring regressions.
Trustworthy LLMs
Posting: the role contributes to explainability and safety improvements. Revision exercise: define a small set of unsafe or hard-to-explain response cases and describe how you would assess both model behavior and the limits of your explanation. The case set is a suggested preparation exercise, not a specified company benchmark.
LLM principles and generation
Posting: research experience with the principles and training methods of leading LLMs, such as GPT or LLaMA, and with text generation or dialogue systems is required. Revision exercise: explain one such model’s training stages and trace how a change in training could affect a multilingual dialogue response.

Skills to revise

Python and PyTorch implementation
Posting: proficiency in Python and PyTorch coding is required. Revision exercise: implement a small training loop with a validation pass, then diagnose a deliberately introduced loss or shape error. The small-scale implementation is a suggested way to rehearse the required coding skills.
Data structures and algorithms
Posting: excellent coding, data-structure, and basic algorithm skills are required. Revision exercise: implement a memory-conscious way to group tokenized examples by length; state its time and space costs and test edge cases. This is practice, not a specified Shopee task.
Model experiments and statistical analysis
Posting: the engineer conducts model-performance experiments, compares emerging technologies, and applies statistical expertise. Revision exercise: write an experiment plan comparing two model variants, including controlled variables, multilingual generation measures, uncertainty, and a decision rule. The measures and rule are preparation suggestions.
Training-problem diagnosis
Posting: the role calls for solving problems in large-model training and applications. Revision exercise: given a hypothetical run with unstable loss, list plausible causes, the observations that would distinguish them, and the order of your checks. Do not treat the hypothetical failure as a reported Shopee issue.
Communicating research trade-offs
Posting: the engineer explores new technologies and works collaboratively with the team. Revision exercise: give a short technical comparison of two pre-training or alignment approaches, separating experimental evidence, unresolved risks, and a proposed next test. This presentation format is a preparation suggestion.

Your preparation checklist

Prepare a research evidence sheet
List relevant doctoral research, LLM training work, and text-generation or dialogue-system research, with dates, methods, and outcomes where available. These correspond to stated requirements; do not imply experience you cannot substantiate.
Build a multilingual pre-training walkthrough
Prepare one project walkthrough, or a clearly labeled hypothetical design, covering data choices, an Instruction Pretraining or Mixture-of-Experts approach, training trade-offs, and how you would diagnose a bottleneck. The Compass team highlights multilingual models, data strategies, and training efficiency.
Work through an alignment design exercise
Sketch how you would compare SFT and RLHF for a multilingual model: training inputs, evaluation criteria, likely failure modes, and a decision rule. Both methods appear in the role’s responsibilities.
Draft a model-evaluation plan
Create a compact experiment table comparing models or methods across language coverage, text quality, explainability, and safety. Specify baselines, statistical checks, and what result would prompt a follow-up experiment.
Refresh implementation examples
Prepare a Python and PyTorch example that shows your approach to data structures, algorithms, and debugging a large-model training problem. Be ready to explain design choices and limitations rather than just present code.
Write questions for the team
Ask how Compass prioritizes multilingual evaluation, which training or inference bottlenecks matter most, and how research comparisons are assessed for business use. These are questions to investigate, not claims about current team processes.

Availability last checked: 4 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in Singapore. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…