New opportunity

Agent Application Algorithm Engineer

Shopee · Singapore

About this role

Shopee’s AI application team develops LLM-powered digital personas intended to act as personal assistants and social bridges. The Agent Application Algorithm Engineer will build agent capabilities and help develop AI-native applications for social interaction and communication.

Responsibilities

Build core agent logic for task planning and orchestration, tool calling, multi-turn dialogue management, memory, RAG, context engineering and multi-agent collaboration.
Lead continuous pre-training and post-training for specific domains and business scenarios. Develop high-quality datasets and data pipelines, design reinforcement-learning reward models, and improve instruction following, reasoning, task completion, role-playing and personalized dialogue, including multimodal conversational experiences.
Develop short- and long-term memory architectures to address forgetting and dispersed attention in long contexts, improving consistency in sustained interactions.
Build multimodal RAG systems and optimize recall, ranking, long-text processing and multi-document synthesis. Develop an agent tool layer integrating external APIs, MCP, search, code interpreters, browsers, sandboxes and third-party services.
Tune prompts and context management for product needs. Design quantitative evaluations, monitor product metrics, use findings to guide agent and model optimization, and explore new AI applications.

Requirements

Master’s degree or above in Artificial Intelligence, Computer Science, Mathematics or a related field; strong programming, problem-solving and engineering skills.
Proficiency in PyTorch and familiarity with distributed training frameworks such as DeepSpeed and Megatron. Experience with LLM training and inference techniques, including data synthesis and filtering, training optimization, prompt engineering, evaluation, deployment and prototyping.
Development experience with agent frameworks such as LangGraph, Google Agent Development Kit or OWL; familiarity with AI development tools such as Cursor and Claude Code. Ability to collaborate across teams from research through deployment.

Preferred

Project experience in agentic reinforcement learning, virtual character generation, multimodal interaction, personality modeling or memory.

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

Skills for this role

AI agentsLLMsPyTorchDeepSpeedMegatronLangGraphGoogle Agent Development KitOWLCursorClaude CodeRAGMultimodal RAGMCPPrompt engineeringContext engineeringReinforcement learningDistributed trainingData synthesisData filteringModel evaluationModel deploymentData pipelinesReward modelingMulti-agent collaborationTool callingDialogue managementMemory architectures

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

Build end-to-end agent behavior
The role builds task planning and orchestration, tool calling, multi-turn dialogue, context handling and multi-agent collaboration, including a tool layer for external APIs and MCP.
Lead domain-specific model improvement
The posting calls for continuous pre-training and post-training, high-quality datasets and pipelines, RL reward models, and improvements to reasoning, task completion and personalized dialogue.
Make sustained interactions consistent
Short- and long-term memory architectures are a stated responsibility, specifically to address forgetting and dispersed attention in long contexts.
Improve multimodal retrieval
The role develops multimodal RAG and optimizes recall, ranking, long-text processing and synthesis across documents.
Tie optimization to product evidence
The posting requires product-tailored prompts and context management, rigorous quantitative evaluations, metric monitoring and using findings to guide agent and model changes.

Likely challenges to think through

Choosing where to intervene
Preparation hypothesis: When an agent fails, changing orchestration, retrieval, prompts or the model may each appear plausible, while each has a different cost. Prepare a diagnostic example that isolates the failure, compares interventions and defines a measurable result.
Balancing memory with relevant context
Preparation hypothesis: Preserving persona and dialogue continuity could compete with keeping long interactions focused on the current task. Prepare to explain what to retain, retrieve or discard, and how to test consistency without sacrificing task completion.
Measuring social dialogue alongside task outcomes
Preparation hypothesis: Personalized, immersive conversation may be harder to evaluate quantitatively than task completion. Prepare a proposed evaluation that separates these aims, uses product-relevant metrics and identifies trade-offs rather than assuming one score captures both.

Practice your answers

Technical practice

How would you design an agent that plans tasks, manages multi-turn dialogue, and coordinates multiple agents?

What to coverSuggested answer focus: explain how you would represent goals, dialogue state, handoffs, and stopping conditions. Discuss when a single agent is simpler than multiple agents and how you would detect orchestration failures.

How would you design short- and long-term memory for a digital persona?

What to coverSuggested answer focus: distinguish current-conversation state from durable user information, and explain how information is selected, updated, and retrieved. Address forgetting, conflicting memories, and consistency across long interactions.

How would you build and improve a multimodal RAG pipeline for long documents?

What to coverSuggested answer focus: walk through recall, ranking, long-text handling, and synthesis across documents and modalities. Explain how you would measure retrieval quality separately from the quality of the final response.

How would you design an agent tool layer for external APIs and MCP integrations?

What to coverSuggested answer focus: describe tool selection, input validation, result handling, and recovery from failures across services such as search or code interpreters. Discuss the trade-off between giving an agent flexible tool access and keeping its actions controlled.

How would you create datasets for continuous pre-training and post-training in a specific domain?

What to coverSuggested answer focus: cover data synthesis, filtering, pipeline quality checks, and the different objectives of the two training stages. Explain how you would detect whether gains on domain tasks come at the expense of broader model behavior.

How would you design a reward model for agent task completion and dialogue quality?

What to coverSuggested answer focus: define observable signals for instruction following, reasoning, task completion, and persona consistency. Discuss ambiguous preferences, reward exploitation, and how you would validate that reward improvements correspond to better interactions.

How would you diagnose and optimize a PyTorch LLM training run using distributed training?

What to coverSuggested answer focus: explain how you would distinguish data, optimization, memory, and throughput bottlenecks. Discuss relevant trade-offs in frameworks such as DeepSpeed or Megatron without assuming a particular infrastructure setup.

How would you adapt context and prompts for different product requirements?

What to coverSuggested answer focus: describe what belongs in the active context, how you would manage context length, and how prompts interact with retrieval and memory. Explain how you would test changes for task success and dialogue consistency rather than relying on illustrative examples alone.

How would you evaluate personalized, role-playing, and multimodal conversations?

What to coverSuggested answer focus: propose quantitative measures for task completion, persona consistency, and conversational quality, including longer interactions. Distinguish what automated checks can capture from what may need carefully designed human assessment.

How would you turn product metrics into an agent optimization plan?

What to coverSuggested answer focus: connect a reported weakness to a measurable hypothesis, a targeted change, and an evaluation plan. Discuss how you would separate model, retrieval, memory, and tool-layer causes before deciding what to optimize.

Behavioural practice

Tell me about a time you took an AI idea from research or prototype through deployment.

What to coverSuggested answer focus: use situation, action, and result to show your own contribution and how you worked with other teams. Explain what evidence led you to proceed, revise the approach, or stop.

Describe a time you investigated a difficult model or system-quality regression.

What to coverSuggested answer focus: give the situation, the diagnostic steps you personally took, and the result. Explain how you narrowed competing causes and verified that the change addressed the problem.

Tell me about a time you had to reconcile different priorities across collaborating teams.

What to coverSuggested answer focus: describe the situation, your actions to clarify trade-offs and reach a decision, and the result. Keep the example grounded in your actual role rather than attributing the whole outcome to yourself.

Describe a time an evaluation result challenged your preferred technical approach.

What to coverSuggested answer focus: set out the initial decision, the evidence that challenged it, what you changed, and the result. Show how you distinguished a meaningful finding from a measurement or data-quality issue.

Tell me about a time you improved the efficiency of an AI development workflow.

What to coverSuggested answer focus: describe the situation, the improvement you implemented, and its observed result. Explain how you checked that faster iteration did not weaken engineering or evaluation quality.

Sharpen your knowledge

Topics to prepare

Agent planning and orchestration
Suggested revision: diagram how an agent plans a task, calls tools, manages multi-turn dialogue and coordinates with another agent; identify where state is retained. The posting explicitly assigns these core agent capabilities.
Continuous training and reward design
Suggested revision: outline a domain-specific pre-training and post-training cycle, including data quality checks and a reward model for instruction following or task completion. The role explicitly leads these efforts for business scenarios.
Short- and long-term memory
Suggested revision: compare what belongs in immediate dialogue context versus persistent memory, then trace a long interaction for forgetting or inconsistent recall. The posting names both memory horizons and long-context attention dispersion.
Multimodal RAG architecture
Suggested revision: sketch a retrieval pipeline and explain how recall, ranking, long-text processing and multi-document synthesis affect the final answer. Each is an explicitly named area of optimization.
Evaluation of persona interactions
Suggested revision: propose quantitative tests for task completion, dialogue consistency and personalized responses, and explain which product metrics would guide iteration. The posting calls for rigorous evaluation and metrics-driven optimization of digital-persona applications.

Skills to revise

PyTorch training and inference
Suggested exercise: implement a small PyTorch training and inference prototype, then explain how you would optimize training and scale it with DeepSpeed or Megatron. PyTorch proficiency, distributed-framework familiarity and LLM training/inference experience are stated requirements.
Dataset and data-pipeline development
Suggested exercise: create a small domain-dialogue dataset, synthesize additional examples and apply documented filtering checks before training. The role explicitly includes high-quality datasets, data pipelines, synthesis and filtering.
Agent framework and tool integration
Suggested exercise: prototype a multi-turn agent in one listed framework—LangGraph, Google Agent Development Kit or OWL—with a tool call and a clearly defined MCP or external-API interface. Framework development and an agent tool layer are explicit in the posting.
RAG implementation and diagnosis
Suggested exercise: prototype retrieval, ranking and synthesis over several documents; inspect cases where the right information was missed or conflicting sources were combined poorly. The posting calls for building and optimizing these RAG modules, including multimodal RAG.
Prompt, context and evaluation iteration
Suggested exercise: create two prompt-and-context variants for a sustained persona dialogue, run the same test set against both, and report quantitative differences and failure cases. Prompt tuning, context management and evaluation aligned with product needs are explicit responsibilities.

Your preparation checklist

Prepare an agent project walkthrough
The role calls for task planning, tool calling and multi-turn dialogue. Select a relevant project, if available, and prepare an architecture diagram showing orchestration, context handling, your decisions and measured results. Label any proposed design as hypothetical.
Practice a memory and retrieval design
Sketch short- and long-term memory for a digital persona, alongside multimodal retrieval with recall, ranking and multi-document synthesis. Explain how you would test forgetting and consistency across sustained conversations.
Map out the agent tool layer
Prepare a design exercise showing how an agent would call an external API or MCP tool, handle failures and return to dialogue. Explain where an agent framework such as LangGraph, Google Agent Development Kit or OWL would fit.
Assemble training and engineering evidence
The posting asks for PyTorch, familiarity with distributed training, and LLM training and inference. Prepare concise examples, where available, covering data synthesis and filtering, pipeline design, training optimization, reward-model design, deployment or prototyping; distinguish direct experience from conceptual knowledge.
Draft a quantitative evaluation plan
Propose measures for instruction following, reasoning, task completion and personalized dialogue, plus an experiment that uses product metrics to guide model or agent changes. State what the measures would and would not show.
Prepare a research-to-deployment account
The role involves cross-team work from research through deployment. Outline one relevant collaboration, if available, with the problem, your contribution, technical trade-offs and how findings informed the next iteration.
Write targeted questions for Shopee
Ask which digital-persona use cases are the team's near-term focus, how it evaluates sustained interactions, and how responsibilities are divided between agent engineering and continuous model training.

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…