New opportunity

Chatbot Algorithm Engineer - Marketplace Intelligence and Data (2027 Graduate)

Shopee · Singapore

About this role

Shopee’s Singapore-based Intelligent Customer Service Chatbot team seeks a Chatbot Algorithm Engineer to build AI assistants and agentic applications for customer service and recommendation. The role sits within Marketplace Intelligence and Data and focuses on production systems that address dialogue, context and multilingual customer-service challenges.

Responsibilities

Design algorithms for AI assistants, including LLM post-training, supervised fine-tuning, Mixture-of-Experts and preference alignment using reinforcement learning or techniques such as DPO and GRPO.
Build LLM-powered recommendation systems for user-intent understanding, retrieval, coarse and fine ranking, sequential modelling, semantic representation learning and generative recommendation.
Develop production-ready applications that connect models to tools, APIs, knowledge bases, memory systems and business workflows.
Explore agent-loop and graph orchestration; planning, reasoning, tool selection and execution; multi-agent collaboration; skill design and lifecycle management; context and memory management; and reflection and feedback mechanisms.
Build reusable algorithm platforms, evaluation frameworks and experimentation pipelines. Establish offline and online evaluations covering task success, response quality, tool-use accuracy, safety, latency, cost, user satisfaction and business impact; conduct failure analysis and design guardrails.
Work with product, business, operations, data and engineering teams from problem definition through deployment, monitoring, root-cause analysis and user-experience improvement. Apply relevant advances in LLMs, agents, recommendation systems and AI engineering to production problems.

Requirements

Bachelor’s degree or higher in Computer Science or a related field; proficiency in at least one language such as Python, C++ or Go, and at least one deep-learning framework such as PyTorch or TensorFlow.
Strong computer-science fundamentals, including data structures, algorithms and machine learning; familiarity with the principles and implementation of LLM, NLP, recommendation and multimodal models; and an interest in LLM technology and real-world problem solving.
Preferred: prior research or internship experience in natural-language Q&A or recommendation technology.

Shopee will prioritize applicants who already have the right to work in Singapore and do not require visa sponsorship.

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

Skills for this role

PythonC++GoPyTorchTensorFlowLarge Language ModelsNatural Language ProcessingDeep LearningMachine LearningSupervised Fine-TuningMixture-of-ExpertsReinforcement LearningDirect Preference OptimizationGroup Relative Policy OptimizationRecommendation SystemsInformation RetrievalLearning to RankSequential ModellingRepresentation LearningGenerative RecommendationAI AgentsMulti-Agent SystemsTool UseGraph OrchestrationKnowledge GraphsEvaluationExperimentationData StructuresAlgorithmsMultimodal AI

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 AI assistant algorithms
The role covers LLM post-training, supervised fine-tuning, Mixture-of-Experts and preference alignment using methods such as DPO and GRPO. Prepare to explain when each approach is useful and how you would evaluate its effect on assistant behavior.
Build intent-aware recommendations
The job spans user-intent understanding, retrieval, coarse and fine ranking, sequential modelling, semantic representations and generative recommendation. Prepare to walk through how these stages fit together, rather than treating recommendation as a single model.
Engineer production agent workflows
The team seeks LLM applications connected to tools, APIs, knowledge bases, memory and business workflows. Agent orchestration, tool selection and execution, and context management are explicit duties; prepare to explain an end-to-end design and its failure handling.
Make evaluation reusable
The role calls for algorithm platforms, experimentation pipelines, and offline and online evaluation across task success, response quality, tool accuracy, safety, latency, cost, satisfaction and business impact. Prepare to describe measurable criteria and a repeatable way to analyze failures.
Carry solutions through deployment
The job involves product, business, operations, data and engineering partners from problem definition through deployment, monitoring and root-cause analysis. Prepare to explain how a research or model improvement would become a monitored customer-service feature.

Likely challenges to think through

Managing multilingual dialogue context
Preparation hypothesis: The team identifies multilingualism and context as core dialogue challenges. Maintaining useful history across languages could improve answers while increasing context length and cost; prepare to discuss context selection, multilingual evaluation cases and how to detect lost or misleading context.
Balancing agent capability with reliability
Preparation hypothesis: Planning, tool calls, memory and multi-agent delegation add capabilities but also introduce more failure points and latency. Prepare to compare a simple workflow with a more autonomous agent, including tool-use checks, guardrails and fallback behavior.
Reconciling quality with production constraints
Preparation hypothesis: A model may score well offline yet fall short on latency, cost, safety or user satisfaction online—all stated evaluation dimensions. Prepare to explain how you would choose trade-offs, monitor outcomes and use failure analysis to decide what to improve next.

Practice your answers

Technical practice

How would you choose between supervised fine-tuning and preference optimization for a customer-service assistant?

What to coverThe role names SFT, reinforcement learning, DPO and GRPO as post-training approaches. Suggested preparation: explain what training signal each uses, how you would evaluate response quality and safety, and when the added complexity of preference optimization might be justified.

What does a Mixture-of-Experts model change about training and serving?

What to coverMixture-of-Experts is explicitly listed in the role. Suggested preparation: explain expert routing and compare potential capacity gains with routing complexity, compute cost and serving latency.

How would you design intent understanding, retrieval and ranking for an LLM-powered recommendation system?

What to coverThe posting specifies user-intent understanding, retrieval, coarse ranking and fine ranking. Suggested preparation: describe what each stage contributes, how candidates move between stages, and the relevance-versus-latency trade-off.

How could user sequences and semantic representations improve recommendations?

What to coverThe role includes user sequential modelling, semantic representation learning and generative recommendation. Suggested preparation: discuss what each could capture, how you would handle limited or changing interaction history, and how you would compare the approaches against simpler recommendations.

How would you connect an LLM assistant to tools, APIs and a knowledge base for a service task?

What to coverThe posting calls for production-ready applications integrated with tools, APIs, knowledge bases and business workflows. Suggested preparation: walk through tool selection, validation of inputs and outputs, handling failed calls, and when the assistant should avoid taking an action.

When would you use a simple agent loop, graph orchestration or multiple agents?

What to coverThe role names agent loops, graph orchestration, planning, tool execution and multi-agent delegation. Suggested preparation: compare how each structure handles task dependencies and recovery, while accounting for coordination overhead, latency and observability.

How would you manage context and memory in a multilingual customer-service dialogue?

What to coverThe team identifies multilingualism and context as dialogue challenges, and the duties include context, memory and knowledge management. Suggested preparation: explain how to retain relevant conversational facts, retrieve knowledge and detect when stale or ambiguous context could produce a wrong answer.

How would you evaluate an assistant offline before testing it online?

What to coverThe posting calls for offline and online evaluation across task success, response quality, tool-use accuracy, safety, latency, cost, satisfaction and business impact. Suggested preparation: propose representative test cases and explain which measures need human review or online evidence rather than relying on a single aggregate score.

How would you investigate a rise in incorrect tool use after deployment?

What to coverThe duties include monitoring, observability, failure analysis, root-cause analysis and guardrail design. Suggested preparation: describe how you would isolate the failure stage, inspect relevant traces and evaluations, and test a mitigation without overlooking safety, latency or cost.

How would you make an assistant experiment reproducible and reusable?

What to coverThe role calls for reusable algorithm platforms, evaluation frameworks and experimentation pipelines, alongside programming and deep-learning framework skills. Suggested preparation: explain how you would organize data, model changes, evaluation results and implementation checks so experiments can be compared and failures reproduced.

Behavioural practice

Tell me about a time you clarified an ambiguous problem with people from different functions.

What to coverThe role involves working with product, business, operations, data and engineering teams from problem definition onward. Suggested preparation: use a real situation, the questions and actions you took to align on a goal, and the resulting decision or outcome.

Describe a time you diagnosed an unexpected failure in a technical project.

What to coverMonitoring, root-cause analysis and continuous improvement are stated duties. Suggested preparation: give a real example with the symptoms, how you tested competing explanations, what you changed and what the result showed.

Tell me about a time user or operational feedback changed your technical approach.

What to coverThe posting connects customer-service development with operations and user-experience improvement. Suggested preparation: describe the original approach, the feedback you received, your response and the observed result without claiming an outcome you did not measure.

Describe a time you evaluated whether a new research idea was practical to implement.

What to coverThe role asks engineers to translate relevant advances in LLMs, agents or recommendation into production solutions. Suggested preparation: use a real project or study, explain how you compared expected benefit with implementation constraints, and state the decision and result.

Tell me about a time you made work easier for others to reuse or evaluate.

What to coverReusable platforms, supporting tools and experimentation pipelines are part of the advertised work. Suggested preparation: describe a real situation, what you documented or built, how others used it and what changed as a result.

Sharpen your knowledge

Topics to prepare

Computer science and machine-learning fundamentals
Review data-structure and algorithm trade-offs, then work through a retrieval problem: compare the time and memory costs of two ways to find relevant items. The posting explicitly requires strong fundamentals in data structures, algorithms and machine learning.
LLM, NLP and multimodal model principles
Explain how an LLM maintains dialogue context, where multilingual customer-service questions can fail, and how text-only and multimodal inputs differ. Review one model architecture at a principles-and-implementation level; the posting explicitly asks for familiarity with LLM, NLP and multimodal models.
LLM post-training and preference alignment
Sketch the data, objective and validation approach for supervised fine-tuning, then contrast DPO, GRPO and reinforcement-learning-based alignment. Explain where Mixture-of-Experts fits. These are named areas of assistant algorithm development in the role.
Recommendation system stages
Trace a customer request through intent understanding, retrieval, coarse ranking and fine ranking; explain where sequential modelling, semantic representations and generative recommendation could contribute. All are named in the job description.
Agent architecture and state management
Draw a small agent loop showing planning, tool selection, execution, feedback and stopping conditions. Add a graph branch, memory boundary and possible task delegation. The role explicitly covers orchestration, multi-agent collaboration, context and skill lifecycle management.
Evaluation and production trade-offs
For a hypothetical customer-service assistant, define offline checks and an online measurement plan covering task success, response quality, tool-use accuracy, safety, latency, cost and user impact. Explain how a failed case would lead to root-cause analysis or a guardrail; these are explicit responsibilities.

Skills to revise

Implement and test algorithms in a stated language
Practice a small retrieval or ranking function in one language named in the requirements—Python, C++ or Go—and test edge cases and complexity. Proficiency in at least one such language is required; this exercise is suggested preparation, not an assigned project.
Use a deep-learning framework
In PyTorch or TensorFlow, implement a small model training loop with validation and inspect a failed prediction. The posting requires skill in at least one of these frameworks and familiarity with model implementation.
Prototype a recommendation pipeline
Using synthetic customer intents and items, build a minimal retrieval stage and two ranking passes; inspect cases where the right item is lost early. This suggested exercise mirrors the posting’s retrieval, coarse-ranking and fine-ranking work.
Connect an assistant to tools and state
Prototype a bounded customer-service flow that chooses between a mock API and a small knowledge base, records necessary dialogue context, and handles a failed tool call. This suggested exercise reflects the role’s explicit tool, API, knowledge-base and memory integration responsibilities.
Build a reproducible evaluation exercise
Create a small set of example dialogues with expected outcomes; score task completion and tool-call correctness, log failures, and propose a guardrail. Then describe what would need online measurement rather than claiming offline scores prove user satisfaction. The posting calls for evaluation frameworks, observability and offline and online evaluation.

Your preparation checklist

Map evidence to the requirements
Prepare a one-page list of relevant coursework, code or projects showing computer-science and machine-learning fundamentals, one programming language, and one deep-learning framework. Note any Q&A or recommendation research or internship work; the posting lists that experience as preferred.
Prepare a project walkthrough
Choose a relevant project, if available, and explain the problem, your contribution, model or system choices, results, and a failure you investigated. The role focuses on customer-service assistants and recommendation systems; label any hypothetical example as such.
Sketch a customer-service agent
As a practice design exercise, diagram how an assistant could use a knowledge base, API tools and memory to resolve a multilingual customer request. Explain tool selection, context handling, escalation or guardrails, and what you would monitor; this is a preparation exercise, not a stated interview task.
Work through a recommendation design
Prepare a second practice design covering intent understanding, retrieval, coarse and fine ranking, and where sequential or semantic representations might help. State how you would test relevance and handle trade-offs rather than assuming every technique is needed.
Compare model-improvement approaches
Make a short comparison of supervised fine-tuning, preference optimization such as DPO or GRPO, and Mixture-of-Experts. For each, be ready to explain its purpose, data needs, implementation considerations and a reason not to use it for a given problem.
Build an evaluation and debugging plan
Draft a small offline test set and an online measurement plan for an assistant. Include task success, response quality, tool-use accuracy, safety, latency and cost, then show how a failed case would lead to root-cause analysis and a proposed fix.
Write questions for the team
Prepare questions about which customer-service or recommendation problems are priorities, how multilingual quality is evaluated, and how product, operations, data and engineering teams work together from deployment through monitoring.

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…