AI research scientist

What does an AI research scientist do?

An AI research scientist designs and runs experiments to build new AI methods or adapt existing ones to business problems.

The job covers model building, testing, and writing up what the results mean.

In Singapore, you find this role in technology firms, banks, government agencies, research institutes and universities. Applied roles usually sit in an AI or data team that reports to a head of AI or a chief data officer.

You work with machine learning engineers, data engineers, product managers, legal and compliance teams, and senior managers who decide which ideas get funded.

What the work involves

• Read recent papers and decide which ideas are worth testing on your organisation's data.
• Write Python and PyTorch code to train, fine-tune and test models, and track each experiment.
• Check that the data is clean, suitable and legally usable before you train on it.
• Build evaluation sets and run tests to see whether a model is accurate, repeatable and explainable.
• Present findings to business stakeholders and senior management, and flag ideas that may be patentable.

Pay in Singapore

Back to top

Skills an AI research scientist needs

Back to top

AI research scientist job description

We are hiring an AI research scientist to run applied research projects in our AI team in Singapore.

You will turn business problems into research questions, build and test models, and report results to stakeholders. You will also help assess whether new solutions can be protected as intellectual property.

Responsibilities

• Review academic and industry research and propose new research topics or use cases for the business.
• Design experiments, simulations and tests for machine learning and large language model (LLM) projects.
• Train and fine-tune models using Python, PyTorch and Hugging Face tools.
• Explore new data sources and judge whether the data is good enough and allowed for research use.
• Build evaluation methods that measure accuracy, fairness, repeatability and traceability of model results.
• Document methods and findings so other teams can reproduce and review them.
• Plan and track timelines for active and upcoming research projects.
• Present outcomes to senior management, business teams and, where approved, public forums.
• Carry out a first check on the patentability of AI solutions and support the filing process with legal teams.
• Work with machine learning engineers to hand over promising prototypes for production.

Requirements

• Master's degree or PhD in computer science, statistics, mathematics, engineering or a related field. Some applied roles accept a bachelor's degree with strong research or project work.
• Typically 0 to 5 years of research or applied machine learning experience, depending on the level.
• Strong Python skills, with working knowledge of PyTorch, Pandas and Jupyter notebooks.
• Solid grounding in statistics and machine learning, including how to design and read experiments.
• Experience with Hugging Face tools and LLM fine-tuning is often asked for, along with a record of papers, patents or open projects.

Back to top

AI research scientist career path

Most AI research scientists start in a junior research or data science role, move into applied research, then lead projects and teams.

The SkillsFuture framework lists the applied researcher as one role, and the other titles below are common Singapore job titles.

1. Research assistant or junior data scientist

   Entry

   You run experiments set by senior colleagues, clean data and reproduce results from papers.

2. AI applied researcher

   Mid · SkillsFuture title

   You design and run AI research projects, find use cases for the business and report results to stakeholders.

3. Senior AI research scientist

   Senior

   You set the research direction for a problem area, review the work of others and advise on patentable ideas.

4. Principal or lead research scientist

   Lead

   You lead a team of researchers, manage project portfolios and represent the work to senior management.

5. Head of AI research

   Head

   You own the research strategy, hiring and budget, and decide where the organisation invests in AI.

Steps marked "SkillsFuture title" follow the career pathway in the SkillsFuture Skills Framework. The others are titles commonly used in Singapore.

Related jobs and their pay

Back to top

AI research scientist interview questions

We wrote these questions from the skills this job needs. They are not an official question bank.

1. A model scores very well on training data but poorly on new data. How do you find the cause and fix it?

   What a good answer covers

   • Names overfitting and checks the gap between training and validation scores
   • Checks for data leakage and mismatched train and test distributions
   • Tries regularisation, simpler models, more data or early stopping
   • Uses cross-validation and learning curves to confirm the fix

2. Model B scores two points higher than Model A on your test set. How do you decide whether the gain is real?

   What a good answer covers

   • Asks about test set size and variance across random seeds
   • Uses confidence intervals, bootstrap or a paired significance test
   • Checks whether the gain holds across data slices
   • Considers whether the difference matters to the business

3. You have 2,000 labelled examples and need to adapt an open-source LLM. How do you approach fine-tuning?

   What a good answer covers

   • Compares prompting and retrieval against fine-tuning before committing
   • Explains parameter-efficient methods such as LoRA against full fine-tuning
   • Splits data properly and watches for overfitting on a small set
   • Compares against the base model on a held-out evaluation set

4. Your PyTorch training loss is not going down. What do you check?

   What a good answer covers

   • Tests the training loop on a tiny batch to see if the model can overfit it
   • Checks learning rate, optimiser, and gradient values for vanishing or exploding gradients
   • Confirms labels, data loading and normalisation are correct
   • Checks that gradients are zeroed and the model is in train mode

5. Describe how you would use Hugging Face to fine-tune and evaluate a text classifier.

   What a good answer covers

   • Loads a pretrained model and matching tokeniser
   • Uses the datasets library and handles padding and truncation
   • Sets up a trainer with metrics suited to class balance, such as F1
   • Saves the model, config and results so the run can be reproduced

6. How would you build an evaluation set for a customer-service chatbot used by a Singapore telco?

   What a good answer covers

   • Samples real queries across topics, including Singlish and mixed-language messages
   • Defines clear scoring criteria for correctness, tone and safety
   • Includes hard and adversarial cases as well as easy ones
   • Checks agreement between human raters and, if used, an LLM judge
   • Keeps the set separate from any data used for tuning

7. A 5 GB CSV file will not load in a Jupyter notebook on your laptop. How do you explore it with Pandas?

   What a good answer covers

   • Reads the file in chunks or selects only the columns needed
   • Sets smaller data types and uses categories for repeated text
   • Samples the data for early exploration
   • Moves to a larger machine or a columnar format when needed

8. How do you make sure another researcher can reproduce your experiment?

   What a good answer covers

   • Fixes random seeds and records library versions
   • Tracks code, data versions, parameters and results in an experiment log
   • Keeps code in version control as well as in notebooks
   • Writes a short note on method and known limits

1. Your new model beats the baseline offline, but the business team sees no gain in a pilot. What do you do?

   What a good answer covers

   • Compares offline data with live data for drift or mismatch
   • Checks that the offline metric reflects what the business cares about
   • Reviews how users actually interact with the model
   • Proposes a better metric or test and reports the finding honestly

2. Senior management wants a generative AI use case for a bank's compliance team and gives you three months. How do you scope the work?

   What a good answer covers

   • Interviews compliance staff to find tasks that are repetitive and well defined
   • Defines success measures and a baseline before building
   • Plans a small prototype with clear stop points
   • Raises data access, explainability and risk concerns early
   • Sets out what you will deliver at each month

3. Two annotators disagree on about a third of the labels in your training data. What do you do?

   What a good answer covers

   • Measures agreement with a statistic such as Cohen's kappa
   • Reviews examples to find unclear label guidelines
   • Rewrites guidelines and relabels a sample
   • Decides how to treat the remaining disagreement, such as adjudication or soft labels

1. Tell us about a research idea that did not work. What did you do next?

   What a good answer covers

   • Describes the hypothesis and how it was tested
   • Explains how and when the team decided to stop
   • Shares what was learned and how it changed later work
   • Shows honesty about the result

2. Describe a time you explained a technical result to people with no machine learning background.

   What a good answer covers

   • Starts from the business question rather than the method
   • Uses simple words and an example or chart
   • Says clearly what the model cannot do
   • Checks that the audience understood and acted on it

1. Your training data includes customer records from Singapore clients. How do you handle it under the Personal Data Protection Act (PDPA)?

   What a good answer covers

   • Confirms with the data protection officer that the purpose is covered by consent or another valid basis
   • Uses only the data needed and removes or masks personal identifiers
   • Controls access and storage of the data
   • Records where the data came from and how it was used

2. You are building a model for a financial institution. What expectations from the Monetary Authority of Singapore (MAS) on AI use should you consider?

   What a good answer covers

   • Mentions the MAS FEAT principles and the newer MAS Guidelines on Artificial Intelligence Risk Management for financial institutions
   • Explains how to test for bias across customer groups
   • Describes documentation and explainability for model decisions
   • Notes that the institution's risk and compliance teams need to review the model

Common questions about an AI research scientist job

Sources