New opportunity

Forward Deployed Engineer, GenAI (Mandarin, English)

Google · Singapore

About this role

Google Cloud seeks a Generative AI Forward Deployed Engineer in Singapore to build and deploy production-grade AI applications with customers. The role bridges Google’s AI products and customer infrastructure, turning prototypes into agentic solutions that deliver measurable return on investment. Google will prioritize applicants who already have the right to work in Singapore and do not require visa sponsorship.

Responsibilities

Code, debug, and jointly launch agentic workflows, including multi-agent systems and MCP servers, within customer environments.
Integrate Google AI products with customer APIs, legacy data, and security perimeters; resolve data-readiness, integration, and state-management barriers to production.
Build evaluation pipelines and observability frameworks to assess accuracy, safety, and latency. Work with customer engineering teams to establish development practices and support adoption.
Turn recurring deployment issues into reusable modules or product feature requests, and explain technical concepts to executives and Mandarin-speaking audiences.

Minimum qualifications:

Bachelor’s degree in Engineering, Computer Science, a related field, or equivalent practical experience; 2 years of software development experience using Python or a similar coding language.
Experience architecting AI systems on cloud platforms such as Google Cloud Platform; building structured and unstructured data pipelines using vector databases and RAG-like architectures; and taking production-grade AI solutions from conception to customer launch.
Fluent Mandarin and English for supporting regional clients.

Preferred qualifications

Master’s or PhD in AI, Computer Science, or a related technical field; experience leading technical discovery sessions.
Experience implementing multi-agent systems with frameworks such as LangGraph, CrewAI, or Agent Development Kit and patterns such as ReAct, self-reflection, or hierarchical delegation.
Knowledge of LLM-native metrics, including tokens per second and cost per request, and techniques for state-management optimization and granular tracing.

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

Skills for this role

PythonGoogle Cloud PlatformVertex AIGeminiGenerative AIAI agentsMulti-agent systemsRetrieval-Augmented GenerationVector databasesData pipelinesLangGraphCrewAIAgent Development KitReActModel Context ProtocolAPI integrationAI evaluationObservabilityLarge Language ModelsState managementTracingMandarinEnglishTechnical discovery

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

Ship agentic systems that deliver measurable value
The role moves prototypes into production-grade workflows, including multi-agent systems and MCP servers, and calls for measurable ROI. Prepare to explain how you would code, debug, launch, and measure an AI application in a customer environment.
Connect AI products to customer infrastructure
The engineer builds integrations with customer APIs, legacy data silos, and security perimeters, while addressing data readiness and state management. Prepare to discuss concrete architecture and debugging decisions for those production blockers.
Evaluate and observe production behavior
The role builds evaluation pipelines and observability frameworks for accuracy, safety, and latency. Prepare to explain how you would detect failures, trace their causes, and use results to guide changes.
Co-build for lasting adoption
The posting calls for working with customer engineering teams on development practices and end-user adoption, and explaining technical concepts to executives and Mandarin-speaking audiences. Prepare to communicate the same technical decision at both engineering and non-technical levels in Mandarin and English.
Turn field lessons into reusable improvements
The engineer identifies recurring friction and converts it into reusable modules or product feature requests. Prepare to distinguish a customer-specific fix from a pattern worth sharing with engineering teams.

Likely challenges to think through

Balancing rapid delivery with enterprise constraints
Likely challenge, not a confirmed employer problem: moving quickly from prototype to launch may conflict with legacy integrations, security perimeters, and data-readiness limits. Prepare to explain how you would identify the blocking dependency, choose a viable scope, and validate the integration before launch.
Managing agent quality against runtime costs
Likely challenge, not a confirmed employer problem: more complex agentic workflows may improve a result while increasing latency, state-management complexity, or cost per request. Prepare to describe evaluation and tracing that make accuracy, safety, latency, and cost trade-offs visible.
Balancing bespoke deployment with repeatability
Likely challenge, not a confirmed employer problem: a solution tailored to one customer may be hard to reuse, while excessive standardization may miss that customer's needs. Prepare to explain when to build a reusable module, when to propose a product feature, and how to support the customer's team in maintaining the deployed solution.

Practice your answers

Technical practice

How would you architect a customer-facing GenAI application using Gemini and Vertex AI?

What to coverSuggested preparation: Explain how the model, application code, customer data and cloud services would fit together. Discuss the trade-off between a quick prototype and an architecture that can be operated in production.

How would you build a RAG pipeline across structured records and unstructured documents?

What to coverSuggested preparation: Cover ingestion, data preparation, retrieval with a vector database and how retrieved content reaches the model. Explain how you would check whether the data is ready and whether retrieval improves answer quality.

When would you use a multi-agent workflow rather than a single agent?

What to coverSuggested preparation: Compare simpler workflows with patterns named in the posting, such as ReAct, self-reflection and hierarchical delegation. Address coordination, added latency and how you would verify that multiple agents improve the result.

How would you design and debug an MCP server used by an agentic workflow?

What to coverSuggested preparation: Describe how you would define the server’s available operations, handle failures and inspect interactions with the agent. Explain how you would keep its access appropriate for a customer environment.

How would you integrate an AI application with customer APIs, legacy data and security perimeters?

What to coverSuggested preparation: Walk through interface discovery, access constraints and testing against live-system behavior. Explain how you would isolate an integration failure before changing application logic.

How would you manage state in a production agentic workflow?

What to coverSuggested preparation: Distinguish information needed for the current step from information that must persist across steps. Discuss how state choices affect correctness, recovery, latency and cost.

How would you evaluate an agentic application before a customer launch?

What to coverSuggested preparation: Outline an evaluation pipeline that tests accuracy, safety and latency on representative tasks. Explain what evidence would support a launch decision and how you would investigate failures.

What observability would you add to diagnose agent behavior and performance?

What to coverSuggested preparation: Discuss granular traces across model calls, retrieval and other workflow steps, alongside latency, tokens per second and cost per request. Explain how those signals help distinguish a quality problem from a performance or cost problem.

How would you move a working prototype to a production launch with measurable ROI?

What to coverSuggested preparation: Identify the operational, evaluation and integration gaps you would resolve with the customer engineering team. Explain how you would define a relevant outcome measure rather than treating a successful demo as proof of value.

How would you troubleshoot a deployment blocked by poor data readiness?

What to coverSuggested preparation: Explain how you would use code and targeted tests to locate issues in source data, pipeline transformations or retrieval. Describe the trade-off between a limited fix for launch and a reusable correction.

Behavioural practice

Tell me about a time you led technical discovery with stakeholders.

What to coverSuggested preparation: Give the situation, the questions and actions you used to uncover the real technical constraint, and the resulting decision. Make clear how you handled uncertainty rather than assuming the initial request defined the solution.

Describe a time you co-built a solution with another engineering team.

What to coverSuggested preparation: Use a situation, action and result to show how you divided work, resolved differences in development practices and supported adoption. Focus on your contribution without claiming sole ownership of a joint launch.

Tell me about a time you explained a complex technical trade-off to a non-technical audience.

What to coverSuggested preparation: Describe the audience and decision, how you adapted the explanation, and the result. If you have a relevant example, address communicating across Mandarin- and English-speaking stakeholders; do not invent one.

Describe a recurring deployment problem you turned into a reusable improvement.

What to coverSuggested preparation: Explain the repeated pattern, the evidence you gathered, and whether your action was a reusable module or a product-feedback proposal. State the result, including any limits on what was adopted.

Tell me about a customer launch where accuracy, safety or latency goals conflicted.

What to coverSuggested preparation: Set out the situation, the trade-off you recommended and the actions you took with stakeholders. Describe the measured result or, if the outcome was unresolved, what the team learned.

Sharpen your knowledge

Topics to prepare

Cloud AI architecture and customer boundaries
Sketch how a Google Cloud AI application would connect to customer APIs, legacy data and security perimeters. Mark where integration and data-readiness issues could block launch. This reflects the required cloud architecture experience and the role’s customer-environment responsibilities.
Enterprise retrieval and data pipelines
Draw a pipeline for structured and unstructured data, showing where a vector database and RAG-like retrieval fit. Explain how you would check whether retrieved material is usable. Both data types, vector databases and RAG-like architectures appear in the minimum qualifications.
Agent orchestration and state
Diagram an agentic workflow, its state transitions and failure points; then compare a single-agent design with a multi-agent design. Multi-agent systems are a responsibility, while frameworks such as LangGraph, CrewAI and ADK and patterns such as ReAct are preferred experience.
Evaluation and observability
Draft an evaluation plan covering accuracy, safety and latency, and explain what granular traces would reveal when a run fails. Evaluation and observability are responsibilities; LLM-native metrics such as tokens per second and cost per request are preferred knowledge.
From prototype to customer launch
Walk through a hypothetical prototype-to-launch decision: identify a production blocker, define a measurable outcome, and decide whether a recurring issue warrants a reusable module or product feature request. This mirrors the stated launch, ROI and field-feedback responsibilities.

Skills to revise

Python integration and debugging
Practice coding a small Python adapter between an AI application and a sample customer API, then diagnose a failed response and document the fix. Python or a similar language is a minimum qualification; coding and debugging integrations are core duties.
Building a retrieval pipeline
With sample structured records and documents, implement ingestion, vector-based retrieval and a RAG-like answer flow; inspect a case where the retrieved evidence is inadequate. This exercises the pipeline experience specified in the minimum qualifications.
Prototyping an agentic workflow
As a preparation exercise, build a small multi-agent workflow or MCP server and test how it handles lost or stale state. The posting names both as examples of production-grade workflows; it does not require a particular agent framework.
Instrumenting an evaluation run
Create a small set of test cases, record accuracy, safety and latency results, and trace one failure through the workflow. If available, calculate tokens per second and cost per request. This follows the evaluation responsibility and preferred metrics knowledge.
Preparing a launch handoff
Draft a concise handoff for a hypothetical customer engineering team: unresolved integration risks, how to reproduce a defect, checks to run before launch, and one recurring issue to turn into a reusable module. This reflects co-building, production launch and field-feedback duties.
Bilingual technical explanation
Practice explaining the same agentic solution and its limitations in Mandarin and English, first to engineers and then to a non-technical executive. Fluency in both languages is a minimum qualification, and translating technical concepts for these audiences is a stated responsibility; leading discovery sessions is preferred experience.

Your preparation checklist

Prepare a customer-launch walkthrough
The role requires experience taking AI solutions from conception to customer launch. If you have a relevant project, prepare a concise walkthrough of your code, debugging decisions, production blockers, launch responsibilities, and measured outcome; distinguish your contributions from the team’s.
Sketch a cloud AI architecture
For practice, draw a hypothetical customer solution using Google Cloud, Vertex AI or Gemini where appropriate. Show how it connects to customer APIs, legacy data, and security boundaries, and explain what you would validate before deployment.
Trace a data-to-answer pipeline
Prepare an example showing how structured and unstructured data would be ingested, checked for readiness, indexed in a vector database, and used in a RAG-like workflow. Identify likely failure points and how you would debug them.
Work through an agentic workflow design
The job includes coding production agentic workflows, such as multi-agent systems and MCP servers. As a practice exercise, diagram agent responsibilities, tool access, state handling, and failure recovery; be ready to explain any framework or pattern you have actually used.
Draft an evaluation and observability plan
Create a sample test-and-monitoring plan for accuracy, safety, and latency, with traces that help isolate failures. Explain how you would assess tokens per second and cost per request where relevant; these are listed as preferred knowledge.
Practice a bilingual customer explanation
The role calls for fluent Mandarin and English and for explaining technical concepts to executives. Prepare a short explanation in both languages of a deployment trade-off, its business impact, and how you would work through it with customer engineers.
Bring focused discovery and feedback questions
Draft questions about a customer’s data readiness, integration constraints, success measures, and engineering ownership. Also prepare an example of how a recurring deployment issue could become a reusable module or a product feature request—an outcome named in the role.

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…