New opportunity

AI Engineer (AI Products)

Nanyang Technological University · Singapore

About this role

AI Singapore’s AI Products Pillar seeks an AI Engineer to develop, train, evaluate and engineer multilingual, multicultural and multimodal foundation large language models, focusing on underrepresented languages and cultures and real-world applications. The role is hosted at Nanyang Technological University’s NTU Main Campus in Singapore under the VP (Artificial Intelligence & Digital Economy)’s office.

Responsibilities

Develop and adapt LLMs through continued pre-training, supervised fine-tuning, instruction tuning and preference optimisation; evaluate model performance.
Build scalable multilingual text pipelines for ingestion, cleaning, filtering, deduplication, tokenisation, contamination detection and dataset mixture creation.
Support distributed training through experiment tracking, checkpointing, monitoring, debugging, profiling and compute optimisation.
Review and reproduce frontier AI research, turning findings into improvements in training, evaluation and deployment. Produce reusable scripts, evaluation tools, model artefacts, APIs, documentation and deployment-ready components.
Collaborate with internal teams, research labs, academic institutions and industry partners. Contribute to technical sharing, workshops, publications and other outreach where relevant.

Requirements

Proficiency in Python and sound software engineering practices, including modular code, Git, testing, documentation and code review. Hands-on experience with PyTorch and NLP/LLM frameworks, particularly Hugging Face Transformers, Datasets, Tokenizers, Accelerate and TRL.
Experience or strong working knowledge of LLM training, tuning, evaluation and adaptation; familiarity with low-resource languages, code-switching, language imbalance, script diversity, tokenisation and multilingual benchmarks.
Ability to build scalable text, dataset and training pipelines; familiarity with distributed and GPU-based training, training stability, checkpointing, experiment tracking and compute optimisation. Familiarity with Linux, Docker, CI/CD, cloud platforms or GPU/HPC clusters.
Ability to assess research papers, design experiments and ablation studies, analyse results and communicate technical trade-offs to varied stakeholders. The team values self-direction, collaboration, reproducibility, reliability and responsible AI development.

Applicants from all disciplines and qualification pathways, including NITEC, Diploma, Bachelor’s, Master’s, PhD or equivalent practical experience, are welcome. A formal qualification in a related technical field is advantageous but not mandatory. Experience with reproducible model workflows, model versioning, automated evaluation and API deployment or serving tools such as vLLM, SGLang or TGI is advantageous.

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

Skills for this role

PythonGitPyTorchHugging Face TransformersHugging Face DatasetsHugging Face TokenizersHugging Face AccelerateTRLLarge Language ModelsMultilingual NLPMultimodal AILLM pre-trainingSupervised fine-tuningInstruction tuningPreference optimisationModel evaluationData pipelinesText processingData deduplicationContamination detectionDistributed trainingGPU trainingExperiment trackingCheckpointingCompute optimisationLinuxDockerCI/CDCloud platformsHPC clusters

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

Develop and evaluate foundation LLMs
The role covers continued pre-training, fine-tuning, instruction tuning, preference optimisation and evaluation of multilingual, multicultural and multimodal models. For preparation, be ready to explain how you would choose an adaptation method and measure its effects.
Build trustworthy multilingual datasets
The posting calls for scalable text ingestion, cleaning, filtering, deduplication, tokenisation, contamination detection and dataset mixture creation, with attention to underrepresented languages and cultures. Prepare to discuss how data decisions affect model evaluation.
Run reliable training at scale
Distributed training support includes tracking experiments, checkpointing, monitoring, debugging, profiling and compute optimisation. Prepare to explain how you would diagnose a training issue while preserving reproducible results.
Turn research into reusable components
The engineer is expected to reproduce relevant research and translate findings into training, evaluation and deployment improvements, including scripts, tools, model artefacts, APIs and documentation. Prepare to distinguish a promising experimental result from a maintainable deliverable.
Communicate across collaborators
The role involves internal teams, research labs, academic institutions and industry partners, plus technical sharing where relevant. The requirements also call for explaining trade-offs and evaluation outcomes to technical and non-technical stakeholders.

Likely challenges to think through

Coverage versus data quality
A likely difficulty is expanding coverage of low-resource languages and code-switching without letting language imbalance, script or tokenisation issues distort results. Prepare a reasoned approach to dataset mixtures and per-language evaluation; this is a preparation hypothesis, not a reported team problem.
Experiment speed versus training reliability
Large-scale distributed training may make it tempting to maximise throughput at the expense of stability or traceability. Prepare to discuss trade-offs among profiling, checkpoint frequency, monitoring and compute use; the posting does not specify the team's infrastructure or bottlenecks.
Research gains versus practical use
A reproduced method may improve a benchmark yet be difficult to maintain or deploy. Prepare to explain how ablations, automated evaluation, documentation and reusable components would inform a decision to adopt it; this is an inferred challenge, not a confirmed employer issue.

Practice your answers

Technical practice

How would you choose between continued pre-training, supervised fine-tuning, instruction tuning and preference optimisation for a multilingual LLM?

What to coverSuggested preparation: Explain what each stage is intended to change and what data and evaluation evidence would justify it. Discuss trade-offs in compute, data quality and preserving performance across languages.

How would you build a scalable pipeline for multilingual training text?

What to coverSuggested preparation: Walk through ingestion, cleaning, filtering, deduplication and dataset mixture creation, including how you would test each stage. Explain how you would keep the resulting dataset reproducible as sources or filters change.

How would you investigate poor performance on a low-resource or code-switched language?

What to coverSuggested preparation: Consider language imbalance, script diversity, tokenisation and the composition of the training mixture. Describe how you would isolate likely causes before changing the tokenizer, data or training approach.

How would you detect and address contamination in a multilingual evaluation dataset?

What to coverSuggested preparation: Distinguish exact duplicates from near-duplicates and translated or reformatted overlaps between training and evaluation data. Explain how you would document exclusions and interpret results if contamination cannot be ruled out.

How would you evaluate whether a model adaptation helps underrepresented languages and cultures?

What to coverSuggested preparation: Propose language-appropriate benchmarks and qualitative checks, and report results by language rather than relying only on an aggregate score. Discuss how you would check for regressions and, where relevant, evaluate multimodal behavior without assuming one metric captures every use case.

A distributed LLM training run becomes unstable. How would you debug it?

What to coverSuggested preparation: Describe how you would use training logs, experiment tracking and checkpoints to locate when the issue began and narrow down data, optimisation or distributed-execution causes. Explain how you would test a fix with a controlled run before resuming at scale.

How would you reduce the compute cost of a large training run without losing useful evidence?

What to coverSuggested preparation: Explain what you would profile and monitor before changing the workflow, including throughput and resource use. Discuss trade-offs among run size, checkpoint frequency, recovery risk and the ability to compare experiments fairly.

How would you decide whether to adopt a method from a frontier LLM paper?

What to coverSuggested preparation: Describe how you would assess the claim, reproduce a meaningful baseline and design ablations that isolate the proposed method's contribution. Explain what results and implementation costs would warrant a change to training, evaluation or deployment.

How would you make LLM training and evaluation code reusable across experiments?

What to coverSuggested preparation: Discuss modular Python and PyTorch code, appropriate use of Hugging Face components, tests, version control and documentation. Explain how you would connect configuration, model artefacts and automated evaluation so another engineer can reproduce a result.

What would you check before turning a model artefact into a deployment-ready API?

What to coverSuggested preparation: Cover model and evaluation versioning, repeatable packaging, testing and operational reliability. Serving frameworks such as vLLM, SGLang and TGI are listed as advantageous experience; discuss how you would compare an appropriate option against the application's needs rather than assuming a required stack.

Behavioural practice

Tell me about a time you collaborated across teams or organisations on a technical problem.

What to coverSuggested preparation: Use a real situation involving differing goals or expertise, and explain your actions to coordinate decisions. State the result and what changed in the working relationship or technical outcome.

Describe a time you explained an uncertain technical result to a non-technical stakeholder.

What to coverSuggested preparation: Set out the situation, how you communicated evidence and trade-offs, and what decision your explanation supported. Include the result without overstating what the evidence proved.

Tell me about a difficult engineering or research setback you worked through independently.

What to coverSuggested preparation: Give the situation, the steps you took to diagnose it or learn what was needed, and when you sought input. Describe the result and how you made the work more reliable or reproducible.

Describe a time responsible AI considerations changed a technical decision.

What to coverSuggested preparation: Use a real example and explain the situation, the concern you identified, the action you took and the result. If relevant, address how you weighed uneven performance across languages or cultures against other goals.

Tell me about a time you shared technical findings so others could use them.

What to coverSuggested preparation: Describe the audience and situation, what you did through documentation, a review or a technical presentation, and the result. Focus on how you made methods, limitations or next steps clear to others.

Sharpen your knowledge

Topics to prepare

LLM adaptation methods
The posting names continued pre-training, supervised fine-tuning, instruction tuning and preference optimisation. Revise by explaining the objective, data and evaluation you would use at each stage, and when you would choose one stage over another.
Multilingual and multicultural model design
The role focuses on underrepresented languages and cultures and names low-resource data, code-switching, language imbalance, script diversity and tokenisation issues. Revise by tracing how each issue could affect data preparation and model performance; propose checks across languages rather than relying on an aggregate score.
Dataset quality and evaluation integrity
The posting calls for filtering, deduplication, contamination detection, dataset mixtures and multilingual benchmark design. Revise by sketching a dataset-to-evaluation workflow and explaining how you would detect overlap with a benchmark and inspect the language mix.
Distributed training trade-offs
GPU-based distributed training, stability, checkpointing, profiling and compute optimisation are stated areas of familiarity. Revise by walking through a training slowdown or instability: what you would monitor, how you would isolate the cause and when you would resume from a checkpoint.
Research reproduction and ablations
The role involves assessing papers, reproducing methods and translating results into practical improvements. Revise by outlining a small ablation for a multilingual LLM method: define a baseline, change one factor, choose evaluation measures and explain the resulting trade-offs.

Skills to revise

Maintainable Python engineering
Python, modular code, Git, tests, documentation and code review are explicit requirements. As a practice exercise, refactor a text-processing script into testable modules, add tests for malformed and multilingual inputs, and write a short review note explaining the changes.
Multilingual text pipeline implementation
The posting requires scalable ingestion, cleaning, filtering, deduplication, tokenisation and dataset mixture creation. Practise building a small pipeline with Hugging Face Datasets and Tokenizers; record per-language counts before and after each step and check for duplicate or contaminated evaluation examples.
PyTorch and Hugging Face tuning
Hands-on PyTorch and Transformers, Accelerate and TRL experience is requested. Practise a small, reproducible fine-tuning run, then document its data, configuration, checkpoint and evaluation results. This is a preparation exercise, not a specified employer project.
Training monitoring and recovery
Experiment tracking, checkpointing, debugging and profiling are named responsibilities. Practise logging training settings and metrics, saving and restoring a checkpoint, and using a profile or monitoring output to explain one compute bottleneck.
Evaluation and technical communication
The posting asks for evaluation tools, ablation studies and clear communication to technical and non-technical stakeholders. Practise comparing two model runs across languages, documenting the evaluation method and limitations, then summarising the trade-off in both a technical note and a short plain-language explanation.
Reproducible deployment workflow
Model versioning, automated evaluation and API deployment or serving tools such as vLLM, SGLang and TGI are described as advantageous, not mandatory. If time permits, practise versioning a model artefact, rerunning its evaluation and documenting how an API would serve it.

Your preparation checklist

Select a truthful project walkthrough
The role spans multilingual LLM development, evaluation and productisation. Prepare a concise walkthrough of a real project, if available, covering your contribution, decisions, results and limitations; otherwise label a proposed design as hypothetical.
Sketch a multilingual data pipeline
The posting specifies ingestion, cleaning, filtering, deduplication, tokenisation, contamination detection and dataset mixtures. Draw a pipeline and explain how you would check low-resource coverage, code-switching, script diversity and data quality.
Design a model-adaptation experiment
The role includes continued pre-training, supervised fine-tuning, instruction tuning and preference optimisation. Prepare a proposed experiment that chooses an approach, states a baseline and explains its compute and data trade-offs.
Build an evaluation plan
The posting calls for multilingual evaluation and research-informed improvements. Draft a small evaluation matrix covering languages or scripts, relevant benchmarks, contamination checks, error analysis and an ablation; explain what conclusions each comparison could support.
Explain training reliability decisions
Prepare a diagram or notes for a distributed GPU training workflow, including experiment tracking, checkpoint recovery, monitoring, debugging, profiling and compute optimisation. Distinguish tools you have used from concepts you would apply.
Gather engineering evidence
The requirements name Python, PyTorch, Hugging Face tools, modular code, Git, tests and documentation. Identify code samples or concrete examples you can discuss, and be ready to explain reproducibility and code-review decisions; API serving is listed as advantageous, not mandatory.
Summarise a paper and its practical implications
The role involves reviewing and reproducing LLM research. Prepare a short summary of one relevant paper, a feasible reproduction or ablation plan, and separate explanations of the findings and limitations for technical and non-technical audiences.
Prepare focused questions for the team
Ask how the AI Products team prioritises underrepresented languages, multimodal work and real-world applications; how it evaluates responsible-AI and deployment trade-offs; and how engineers collaborate with research and industry partners.

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…