New opportunity

LLM Algorithm Engineer (Post Training)

Shopee · Singapore

About this role

Shopee is seeking an LLM Algorithm Engineer (Post Training) in Singapore for its Engineering and Technology team. The role supports post-training of large language models for agent and tool-use applications, including work on multilingual AI applications.

Responsibilities

Develop and deploy post-training pipelines using SFT, DPO, PPO and Reward Modeling. Execute training, hyperparameter tuning, evaluation and online regression.
Optimize agent and tool-use training to improve tool-calling accuracy, multi-turn instruction following and hallucination control. Iterate on models for routing, query rewriting, agent tool calling, web-search decisions, memory and e-commerce search relevance.
Manage post-training data collection, cleaning, synthesis and annotation; help develop annotation guidelines and coordinate with vendors.
Build and maintain automated evaluation combining human review with LLM-as-Judge, produce reproducible effectiveness reports, and experiment with relevant emerging methods.

Requirements

Master's degree in Computer Science, AI or a related field. Prior development experience in at least one LLM post-training area, such as SFT, DPO, PPO, RLHF or Reward Modeling, with the ability to independently carry out key steps from data through training and evaluation.
Experience with at least one mainstream training framework, such as Megatron or veRL, and an understanding of distributed-training fundamentals. Understand agents, tool use and multi-turn dialogue, including data construction and basic alignment approaches for Function Calling.
Strong engineering and troubleshooting skills, including the ability to balance model effectiveness against iteration efficiency. Prior end-to-end post-training deployment experience or participation in 1000-GPU-scale distributed training.

Preferred

Experience with leading large-model teams, or strong understanding or hands-on experience in Agent RL, self-play, synthetic data or inference-time compute.

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

Skills for this role

Large Language ModelsSupervised Fine-TuningDirect Preference OptimizationProximal Policy OptimizationRLHFReward ModelingHyperparameter TuningAgent Tool UseFunction CallingMulti-turn DialogueQuery RewritingWeb SearchE-commerce Search RelevanceData AnnotationSynthetic DataLLM-as-JudgeMegatronveRLDistributed TrainingAgent RLSelf-playInference-Time ComputeTroubleshooting

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

Execute end-to-end post-training
The role covers developing and deploying SFT, DPO, PPO and reward-modeling pipelines, from training and hyperparameter tuning through evaluation and online regression. It calls for independent execution of key steps from data to evaluation.
Improve agent and tool-use behavior
Training optimization targets tool-calling accuracy, multi-turn instruction following and hallucination control. The role also iterates on routing, query rewriting, web-search decisions, memory and e-commerce search relevance.
Own post-training data workflows
The job includes collecting, cleaning, synthesizing and annotating data, plus contributing to annotation guidelines and coordinating with vendors.
Make evaluation reproducible
The role builds automated evaluation that combines human review with LLM-as-Judge and produces reproducible effectiveness reports.

Likely challenges to think through

Balance model gains against iteration speed
Likely preparation challenge, not a confirmed team problem: more tuning or larger experiments may improve effectiveness while slowing iteration. Prepare a concrete way to choose experiments, troubleshoot distributed training and decide when a gain justifies its cost.
Trust evaluation across agent behaviors
Likely preparation challenge: an aggregate score may conceal failures in tool calls, multi-turn behavior or hallucination control, while human and LLM-as-Judge assessments may differ. Prepare an evaluation design with scenario-level checks, human review and reproducible regression comparisons.
Turn varied data into reliable training signals
Likely preparation challenge: inconsistent annotations or synthetic examples could obscure whether changes help across the listed scenarios. Prepare an approach to data cleaning, annotation guidance and quality checks before comparing training runs.

Practice your answers

Technical practice

How would you choose among SFT, DPO, PPO and reward modeling for a post-training objective?

What to coverThe posting names these approaches as parts of its post-training work. Suggested preparation: explain the data and feedback each approach needs, the behavior it can target, and the trade-off between potential gains and training complexity.

How would you tune and troubleshoot a distributed post-training run?

What to coverThe role calls for hyperparameter tuning, distributed-training fundamentals and experience with a framework such as Megatron or veRL. Suggested preparation: discuss how you would diagnose training instability or throughput problems while preserving reproducibility and iteration speed.

How would you construct post-training data for multilingual applications?

What to coverThe page describes a multilingual AI ecosystem and work on data collection, cleaning, synthesis and annotation. Suggested preparation: discuss coverage, data quality, annotation consistency and how you would check whether synthetic examples help rather than distort evaluation.

How would you improve function-calling accuracy in an agent?

What to coverThe role explicitly targets agent tool use and calls for familiarity with function-calling data construction and alignment. Suggested preparation: describe examples of valid calls, arguments and tool responses, then explain how you would measure wrong-tool, wrong-argument and unnecessary-call errors.

How would you evaluate multi-turn instruction following and hallucination control?

What to coverBoth are stated training objectives. Suggested preparation: distinguish failures in retaining instructions across turns from unsupported claims, and explain how test conversations and human review could reveal each.

How would you assess a change to routing, query rewriting or search decisions?

What to coverThe posting identifies routers, query rewriting, web-search decisions and e-commerce search relevance as model-iteration areas. Suggested preparation: explain how you would compare downstream relevance and decision quality against latency or unnecessary tool use, rather than judging rewritten queries alone.

What would you test when iterating on an agent's memory behavior?

What to coverMemory is listed among the core scenarios for model iteration. Suggested preparation: propose tests for using relevant prior context, ignoring irrelevant context and avoiding unsupported recall; treat the specific memory design as unknown.

How would you build a reproducible evaluation using human review and LLM-as-Judge?

What to coverThe role includes automated evaluation that combines both methods and produces reproducible effectiveness reports. Suggested preparation: cover stable test sets, review criteria, judge calibration against humans, disagreement analysis and recording model and evaluation versions.

How would you detect regressions after deploying a post-trained model?

What to coverThe stated pipeline extends through deployment, evaluation and online regression. Suggested preparation: explain how you would compare new and prior behavior on the role's agent and search objectives, investigate regressions and decide whether a gain warrants its operational cost.

How would you evaluate an emerging post-training method before adopting it?

What to coverThe posting asks for experiments with frontier methods; Agent RL, self-play and inference-time compute appear as preferred areas. Suggested preparation: choose one as an example and describe a controlled comparison, expected benefit, compute or iteration cost, and criteria for stopping the experiment.

Behavioural practice

Tell me about a time you took a model change from data through evaluation.

What to coverThe role calls for independently completing key post-training steps. Suggested preparation: use a real situation, specify your actions at each stage and give the observed result without claiming responsibility for steps you did not own.

Describe a difficult training or engineering issue you diagnosed.

What to coverTroubleshooting and balancing effectiveness with iteration efficiency are explicit requirements. Suggested preparation: explain the situation, the evidence and actions that narrowed the cause, the trade-off you made and the result.

Tell me about a time you improved an annotation or data-quality workflow.

What to coverThe work includes annotation guidelines and vendor coordination. Suggested preparation: describe a real quality or consistency problem, your actions with reviewers or vendors, and the resulting change in data quality or workflow.

Describe a time evaluation evidence challenged a proposed model improvement.

What to coverThe role combines human review, LLM-as-Judge and reproducible reporting. Suggested preparation: give the situation, how you investigated conflicting evidence, what you communicated or changed, and the result.

Tell me about a time you had to prioritize model quality against delivery speed.

What to coverThe posting explicitly calls for trade-offs between effectiveness and iteration efficiency. Suggested preparation: describe the constraints, the actions and reasoning behind your choice, and what the outcome showed.

Sharpen your knowledge

Topics to prepare

Post-training methods and objectives
The role names SFT, DPO, PPO and reward modeling, and requires experience in at least one post-training area. Revise by explaining the data, objective and evaluation approach for each method, then justify which you would try first for a specific model-behavior problem.
Function-calling alignment
The posting requires an understanding of agents, tool use, multi-turn dialogue, and function-calling data construction and alignment. Revise by tracing a multi-turn request through tool selection, arguments, tool response and final answer; identify where calling accuracy or hallucination could fail.
Distributed-training fundamentals
The role requires basic distributed-training understanding and experience with a mainstream framework; Megatron and veRL are examples, not both required. Revise by explaining how parallelism, communication and memory constraints affect a post-training run, and how you would diagnose a slowdown.
Evaluation and online regression
The job calls for automated evaluation combining human review and LLM-as-Judge, reproducible reports, and online regression. Revise by outlining how you would compare two model versions on tool-calling accuracy, multi-turn instruction following and hallucination control while checking for regressions.
Effectiveness versus iteration efficiency
The requirements explicitly call for engineering trade-offs between model effectiveness and iteration efficiency. As interview preparation, compare two hypothetical training or evaluation plans and explain what evidence would justify the added cost of the slower plan.

Skills to revise

Execute a post-training run end to end
The role covers data through training, hyperparameter tuning and evaluation. Practice drafting a run plan for one post-training method you know: specify input data, tuning choices, checkpoints, evaluation and a reproducible record of results.
Prepare and audit training data
The posting covers collection, cleaning, synthesis, annotation guidelines and vendor coordination. Practice writing a short annotation guideline for tool-use examples, then audit a small sample for ambiguous labels, duplicates and inconsistent tool responses.
Test agent and search decisions
The job names routing, query rewriting, agent tool calling, web-search decisions, memory and e-commerce search relevance. Practice creating a small set of multi-turn cases and marking the expected decision, tool call and failure mode for each. These cases are a preparation exercise, not a stated employer test.
Build a reproducible evaluation report
The role calls for human review alongside LLM-as-Judge. Practice producing a compact comparison of two hypothetical model versions that records test-set criteria, judge instructions, human-review disagreements and changes in the stated agent metrics.
Troubleshoot a training regression
Strong troubleshooting and end-to-end post-training deployment or large-scale distributed-training experience are listed requirements. Practice walking through a hypothetical run with degraded tool-calling accuracy: isolate possible data, training and evaluation causes, then propose the next diagnostic experiment.

Your preparation checklist

Map the stated requirements to evidence
Prepare a one-page evidence map covering a relevant master’s degree, work in at least one post-training method, a training framework such as Megatron or veRL, and either end-to-end deployment or participation in 1000-GPU-scale training. Mark any gaps rather than implying experience.
Rehearse an end-to-end post-training walkthrough
Select a project you can discuss and explain your role from data preparation through SFT, DPO, PPO, or reward modeling, hyperparameter choices, evaluation, and any online regression work. Bring reproducible results where available.
Explain framework and distributed-training decisions
Prepare a technical account of a framework you have used, distributed-training fundamentals, a troubleshooting incident, and how you balanced model effectiveness against iteration speed.
Practice an agent and tool-use design exercise
Sketch a function-calling, multi-turn workflow for one stated scenario, such as routing or e-commerce search. Define training examples and checks for tool-calling accuracy, instruction following, and hallucination control; present it as a proposed design, not Shopee’s implementation.
Show how you would manage post-training data
Draft a short data plan covering collection, cleaning, synthesis, annotation guidelines, and quality checks. Include how you would resolve ambiguous labels and coordinate annotation work with vendors.
Prepare a reproducible evaluation plan
Outline an evaluation report that combines automated checks, human review, and LLM-as-Judge. Specify baselines, failure categories, regression checks, and what another engineer would need to reproduce the findings.
Bring focused questions for the team
Ask which agent or tool-use scenarios are priorities, how multilingual performance is evaluated, what online regression measures matter, and how the team weighs effectiveness against iteration efficiency.

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…