Read this before anything else. On 22 April 2026 at Cloud Next '26, Google renamed Vertex AI to the Gemini Enterprise Agent Platform and absorbed Agentspace into a unified Gemini Enterprise product. Google has stated that all future roadmap evolution ships under the Agent Platform brand rather than as standalone Vertex AI.
Every GCP AI interview list currently online still says "Vertex AI." If you can explain the rebrand, what carries over, and what requires the new Agent Platform plan, you will be ahead of nearly every other candidate in the room — including, quite often, the interviewer.
Built from placement feedback across Cloud Soft Solutions' APEX batches — questions candidates actually faced at product companies, GCCs, consultancies and startups across Hyderabad, Bengaluru and Pune, plus what our trainers ask when screening. Answers are written the way you should say them: direct first, detail second.
Part 1 — What changed in 2026 (read this first)
1. Vertex AI is now the Gemini Enterprise Agent Platform
Announced and generally available on 22 April 2026 at Cloud Next '26 in Las Vegas. Google describes it as the evolution of Vertex AI, built around four pillars — build, scale, govern, optimize. Existing customers see the new brand appear in their console with no manual migration, and existing APIs remain backward compatible.
The nuance that scores: agents built on the previous generation keep working, but accessing Agent Runtime, Memory Bank, Agent Registry and the optimisation suite requires enabling the new Agent Platform plan. Backward compatible is not the same as automatically upgraded.
2. Agentspace was absorbed into Gemini Enterprise
The standalone Agentspace product is gone as a separate thing, consolidated into Gemini Enterprise alongside the Agent Platform and an Agent Marketplace carrying partner agents from Adobe, Box, Salesforce, ServiceNow and Workday.
3. ADK moved to a graph-based framework
The Agent Development Kit now uses a graph-based framework organising agents into networks of sub-agents, so multi-agent logic is defined explicitly rather than emerging from conversation. ADK v1.0 is stable across four languages. Alongside it, Agent Studio provides a low-code visual path from prompt to deployed agent.
4. The agent control plane is now a real product surface
New components: Agent Registry (catalogue of agents), Agent Identity (agents as first-class principals), Agent Gateway (controlled ingress/egress), Agent Observability, Memory Bank (persistent memory) and Agent-to-Agent Orchestration. If you are interviewing for a platform or architect role, this is the vocabulary to have.
5. A2A v1.0 is in production; MCP servers are managed
Agent2Agent (A2A) protocol reached v1.0 and is reported in production at around 150 organisations. Google also announced managed MCP servers with Apigee acting as an API-to-agent bridge — turning an existing API estate into agent tools without rewriting it. Knowing both A2A and MCP, and how they differ, is close to mandatory for agent roles.
6. Models: Gemini 3.1, Gemma 4, and 200+ in Model Garden
Model Garden now carries 200+ models including Anthropic's Claude. Recent Google models include Gemini 3.1 Pro, Gemini 3.1 Flash Image, Lyria 3 for audio, and Gemma 4 — open-weight under Apache 2.0, built from the same research as Gemini 3, for teams that must self-host.
7. Workspace Studio and Project Mariner
Workspace Studio lets business users build agents across Gmail, Docs, Sheets, Drive, Meet and Chat from plain-language descriptions. Project Mariner is the web-browsing agent. Both come up in "how would non-developers use this" questions.
Part 2 — Fundamentals (0–2 years)
1. What is the Gemini Enterprise Agent Platform?
Google Cloud's unified platform for building, scaling, governing and optimising AI applications and agents — the evolution of Vertex AI, generally available since 22 April 2026. It brings model selection, model building, agent building, integration, DevOps, orchestration and security under one control plane.
2. What happened to Vertex AI?
It was renamed. The services are the same and existing APIs stay backward compatible, so nothing breaks and no migration is required. What changes is that future roadmap work ships under the Agent Platform brand, and the new agent-lifecycle features — Agent Runtime, Memory Bank, Agent Registry, the optimisation suite — require enabling the Agent Platform plan.
What's really being tested: whether you follow the platform. Say both names, say the date, and say what carries over. Candidates who only know "Vertex AI" sound like they stopped reading in 2024.3. What is the difference between AI, ML, deep learning and generative AI?
AI is the broad field. ML is the subset that learns patterns from data rather than following coded rules. Deep learning is ML using multi-layer neural networks. Generative AI is the class of models producing new content — text, image, audio, code — rather than only classifying or predicting.
4. What are the pre-built AI APIs on Google Cloud?
Cloud Vision, Video Intelligence, Speech-to-Text, Text-to-Speech, Translation, Natural Language, and Document AI for structured document extraction. They solve standard tasks with no training. Increasingly these overlap with Gemini's multimodal capabilities — the judgement question is which to reach for.
5. When would you use Document AI instead of sending a scan to Gemini?
When you need deterministic, structured extraction with bounding-box coordinates, high volume at low cost, and prebuilt processors for standard forms — invoices, receipts, IDs, lending documents. Gemini wins on unusual layouts, reasoning about content and multi-document synthesis. Saying "I'd check whether a processor already exists" scores well; jumping straight to an LLM does not.
6. What is a token, and why does it matter for Indian-language applications?
The unit a model reads and writes — roughly four characters in English. Pricing, context limits and latency are measured in tokens. Devanagari, Telugu and other Indic scripts tokenise less efficiently than English, so the same sentence can cost noticeably more. Worth raising unprompted if the role involves Indian-language products.
7. What is a context window, and what is Gemini's?
The total tokens the model considers in one request — system instruction, history, retrieved documents, tool schemas and the response. Gemini has led on long context, with 1M-token windows on Pro-tier models and expansion beyond that in the 3.x line. Long context is a budget to manage, not a reason to stop doing retrieval.
8. What is an embedding?
A vector representing semantic meaning, so similar meanings sit close together in vector space. On Google Cloud you generate them with the text-embedding models and store them in Vector Search, AlloyDB, Cloud SQL with pgvector, BigQuery or Firestore.
9. Temperature and top_p — what do they do?
Temperature scales randomness in token selection: 0 for extraction and classification, higher for creative generation. top_p limits selection to the smallest token set whose cumulative probability exceeds p. Gemini also exposes top_k. Tune one, not all three — otherwise behaviour becomes impossible to reason about.
10. What is grounding, and what grounding sources does Google offer?
Constraining output to verifiable sources. Google offers Grounding with Google Search (public web, with search entry points you are required to display), grounding on your own data via Vertex AI Search or RAG Engine, and grounding on third-party datasets. Google Search grounding is a genuine differentiator against other clouds.
11. What is RAG?
Retrieval-Augmented Generation — retrieve relevant content from your own corpus at query time and pass it as context, so answers are grounded in your sources rather than training data. It addresses freshness, private data and citability in one pattern.
12. RAG or fine-tuning — how do you choose?
RAG for knowledge that changes, is private, or must be cited. Tuning for behaviour — format, tone, domain style, shorter prompts. They combine rather than compete, and most problems need neither because a better prompt solves them.
The trap: proposing fine-tuning to "teach the model our documents." Tuning teaches style, not recall. Say so plainly.13. What is Gemma, and when would you use it over Gemini?
Gemma is Google's open-weight model family — Gemma 4 is Apache 2.0 licensed and built from the same research as Gemini 3. Use it when you must self-host: air-gapped environments, strict data residency, on-device or edge deployment, or when you want to fine-tune weights you control. Gemini for managed capability, Gemma for control.
14. What is an AI agent, as distinct from a chatbot?
An agent receives a goal and decides which tools to call, in what order, across multiple turns, using memory and state, until the goal is met. A chatbot responds. The operationally significant differences are tool access, autonomy and state — and all three are what make agents hard to secure and evaluate.
15. What is BigQuery ML?
Training and running ML models using SQL inside BigQuery, without moving data. It covers regression, classification, clustering, time series, matrix factorisation, and — importantly for 2026 — calling Gemini directly from SQL via remote models for generation, embedding and understanding over tables. For teams whose data already lives in BigQuery this is often the shortest path to production.
16. What is Dialogflow CX and where does it fit now?
Google's conversational agent builder with state-machine-based flows, used heavily in contact centres alongside Contact Center AI. It remains appropriate where you need deterministic, auditable conversation paths — regulated flows, IVR. Generative agents on the Agent Platform suit open-ended tasks. Many production systems use both, with Dialogflow handling the controlled path.
17. What is multimodality, and how does Gemini handle it?
Gemini is natively multimodal — text, image, audio, video and PDF in a single request, rather than separate models stitched together. Practically, this means you can pass a video and ask questions about it without transcribing first, which materially changes how you design media pipelines.
18. What is function calling?
Supplying function declarations so the model returns a structured request to invoke one, which your code executes and returns. The model decides whether and which; your application always controls execution. That boundary is a security decision as much as a design one.
19. What is a hallucination and why does it happen?
Confident output not supported by fact or source, because the model predicts plausible continuations rather than retrieving verified facts. Mitigation is grounding, instructing the model to decline when sources are insufficient, requiring citations, and measuring groundedness as a metric rather than assuming it.
20. What is Gemini Code Assist?
Google's AI coding assistant across IDEs and the Cloud console, with codebase awareness for enterprise repositories. Relevant in interviews mostly as a developer-productivity question — and worth knowing that code assistants are now a governance topic, not just a tooling one.
Part 3 — Gemini Enterprise Agent Platform
21. What are the four pillars of the Agent Platform?
Build — ADK, Agent Studio, Model Garden. Scale — Agent Engine runtime, Memory Bank. Govern — Agent Registry, Agent Identity, Agent Gateway, Agent Observability. Optimize — evaluation, tuning and the optimisation suite. Framing your answer around these four verbs signals you have read the announcement rather than a summary of it.
22. What is Agent Engine?
The managed runtime for deploying agents — it handles scaling, sessions, state and integration so you are not operating your own serving infrastructure. Billed on runtime consumption (vCPU-hours and GB-hours) plus session and memory storage per thousand events.
23. What is Memory Bank?
Managed persistent memory for agents across sessions — user preferences, prior outcomes, learned context. The design questions it raises are retention and deletion policy, because agent memory is personal data subject to the same rights as any other store.
24. What are Agent Registry, Agent Identity and Agent Gateway?
Registry is the catalogue of agents in the organisation — what exists, who owns it, what version. Identity makes an agent a first-class principal with its own credentials and permissions, so you can audit what the agent did rather than what a shared service account did. Gateway controls what an agent can reach. Together they are the answer to "how do you stop agent sprawl."
25. Agent Studio versus ADK — when do you use each?
Agent Studio is the low-code visual builder: fast, good for business users and for prototyping, weaker for complex logic and version control. ADK is code-first with a graph-based framework, source control, testing and CI. Prototype in Studio, productionise in ADK — and say that explicitly, because "which would you use" is testing whether you understand the handoff.
26. What is Gemini Enterprise?
The unified enterprise product that absorbed Agentspace — a front door where employees access agents, combined with the Agent Platform for building them and an Agent Marketplace for partner-built agents from vendors like Salesforce, ServiceNow, Workday, Box and Adobe.
27. How is a Google Cloud AI project structured for production?
Separate projects per environment under a folder hierarchy, with org policy constraining regions and models. Service accounts with least-privilege IAM, Workload Identity Federation instead of keys, VPC Service Controls around the AI perimeter, Private Service Connect for private access, CMEK where policy requires, and Cloud Logging plus Cloud Monitoring wired in from day one.
28. What are VPC Service Controls and why do they matter for AI?
A service perimeter preventing data exfiltration from Google-managed services, even by an identity that holds valid credentials. For AI workloads it is how you stop a compromised or misconfigured job from sending your corpus to a model endpoint outside the perimeter. It is the control regulated Indian clients ask about first.
29. How do you handle data residency for an Indian client?
Deploy in asia-south1 (Mumbai) or asia-south2 (Delhi), and check model availability there — the newest models often reach those regions later, so model choice may be constrained. Use regional rather than global endpoints. Add VPC Service Controls, CMEK and org policy restricting resource locations. Surface the model-availability trade-off in week one, not month six.
30. How do you promote an agent from prototype to production?
Separate projects per environment, infrastructure as code with Terraform, ADK agent definitions and prompts in Git with review, an evaluation suite as a CI gate, deployment to Agent Engine through Cloud Build, registration in Agent Registry with a named owner, tracing to Cloud Trace and Cloud Logging, and a canary rollout comparing metrics. The evaluation gate is the part that distinguishes an engineer from someone shipping prompts by hand.
Part 4 — Gemini models & Model Garden
31. What is Model Garden?
The catalogue of models available on the platform — 200+ as of Cloud Next '26 — spanning Google's own Gemini and Gemma, Anthropic Claude, Llama, Mistral, and specialised open models. Deployment options differ: managed API for first-party and partner models, or self-deployed endpoints for open weights.
32. How do you choose between Gemini Pro and Flash tiers?
Pro for complex reasoning, long-context synthesis and difficult code. Flash for high-volume, latency-sensitive work — classification, extraction, routing, summarisation — at a fraction of the cost. Flash-Lite where volume dominates entirely. The mature answer: run everything on the cheapest tier that passes your evaluation set, and escalate only the requests that fail.
33. What is context caching and when does it pay off?
Caching a large stable prefix — a document corpus, a long system instruction, a codebase — so repeated requests against it are billed and processed at a reduced rate. It pays off when many queries share the same large context. Structure prompts with stable content first and variable content last to make it work.
34. What is Provisioned Throughput?
Reserved model capacity with predictable throughput, rather than competing for shared on-demand capacity. Worth it when volume is high and steady or latency variance is unacceptable. Below the crossover point, pay-as-you-go is cheaper — and saying "I'd model the crossover before committing" is the answer.
35. What is Batch Prediction?
Asynchronous processing of large input sets at reduced cost, reading from and writing to Cloud Storage or BigQuery. Right for bulk classification, offline enrichment, embedding generation and evaluation runs — anything without a user waiting.
36. What are safety settings on Gemini?
Configurable thresholds across harassment, hate speech, sexually explicit and dangerous content categories, applied to both prompt and response, with block levels from none to low-and-above. Some categories cannot be disabled. Always check finishReason and safety ratings in the response rather than assuming text came back.
37. What is structured output on Gemini?
Constraining the response to a supplied JSON schema via responseMimeType and responseSchema, so downstream parsing does not fail. Always prefer schema enforcement over asking for JSON in the prompt and repairing it with regex afterwards.
38. What is a global endpoint versus a regional endpoint?
Global routing gives better availability and higher effective capacity by serving from wherever there is headroom; regional endpoints keep processing within a named region for residency and latency predictability. Regulated workloads take regional even at a cost premium.
39. What is Lyria, and what is Veo?
Lyria is Google's music generation model line (Lyria 3 as of 2026); Veo is video generation; Imagen is image generation. All carry SynthID watermarking. Media roles will probe these; general AI roles usually just want to know they exist and are watermarked.
40. How do you handle model deprecation?
Pin model versions explicitly rather than using floating aliases in production. Track deprecation notices. Maintain an evaluation set so you can qualify a successor model quickly and objectively rather than eyeballing outputs. Model churn is a standing operational cost on every cloud, and interviewers ask because most teams handle it badly.
Part 5 — Prompting & context
41. What belongs in a system instruction?
Role, task, explicit constraints and refusals, output format, tone, and behaviour when information is missing. Keep instructions separate from user data with clear delimiters — merging them is how prompt injection gets in.
42. Zero-shot, few-shot, chain-of-thought — when does each apply?
Zero-shot for well-understood tasks. Few-shot when format or edge cases need demonstrating — three to five diverse examples beat twenty similar ones. Chain-of-thought for multi-step reasoning, though with thinking-enabled models you generally should not prompt for it explicitly.
43. What is a thinking budget?
A control on how much inference-time reasoning a model performs before answering. Higher budgets improve hard reasoning tasks at the cost of latency and tokens; lower or zero suits simple extraction. Tuning it per endpoint rather than globally is the production pattern.
44. What is prompt injection, and how is indirect injection different?
Direct injection is a user instructing the model to ignore its instructions. Indirect injection hides instructions in retrieved content — a document, a web page, an email — so the attack arrives through your RAG pipeline. Indirect is harder and agents make it worse, because agents act on what they read.
Strong answer: "Treat all retrieved content as untrusted, never let an agent's effective privilege exceed the trust level of the least-trusted content in its context, and require human approval for irreversible actions."45. How do you manage a conversation that outgrows the context window?
Summarise older turns into a running summary, keep recent turns verbatim, store full history externally and retrieve selectively, and keep tool schemas out of context where the platform supports selective tool retrieval. Truncating from the front blindly drops the system instruction — a common and silent bug.
46. Long context is 1M+ tokens — why still do RAG?
Cost, latency and accuracy. Filling a million tokens per request is expensive and slow, and attention degrades across very long contexts, so relevant content buried in the middle gets missed. RAG retrieves the few thousand tokens that matter. Long context complements retrieval; it does not replace it.
47. How do you version and test prompts?
Prompts in source control, a golden dataset with expected characteristics, automated evaluation on every change, canary rollout with metric comparison. Treat a prompt change like a code change — it has the same blast radius and none of the type safety.
Part 6 — RAG, Vertex AI Search & Vector Search
48. Walk me through a RAG pipeline end to end.
Ingest → parse → chunk → embed → index → (query) embed query → retrieve top-k → rerank → assemble prompt with citations → generate → evaluate. The follow-up is always which stage you tune first. Answer: retrieval. A generation problem is usually a retrieval problem in disguise.
49. What is Vertex AI Search, and what does it give you out of the box?
A managed search and RAG service — ingestion connectors, layout-aware parsing (PDF, HTML, DOCX, PPTX, XLSX), chunking, embedding, hybrid retrieval, reranking and grounded answer generation, with industry variants for retail, media and healthcare. It removes most of the pipeline you would otherwise hand-build.
50. What is the layout parser?
Document parsing that preserves structure — headings, tables, sections — so chunks align to meaningful units rather than arbitrary character counts. It supports PDF, HTML, DOCX, PPTX and XLSX. It is usually the highest-leverage improvement on a document-heavy corpus.
51. What is Vector Search?
Google's managed vector database (formerly Matching Engine), built on the ScaNN algorithm for approximate nearest neighbour at very large scale with low latency. Use it when you need raw vector retrieval under your own control; use Vertex AI Search when you want the whole pipeline managed.
52. When do you use AlloyDB or BigQuery for vectors instead?
When the data already lives there and joins matter. AlloyDB and Cloud SQL with pgvector let you filter on relational predicates and vectors in one query. BigQuery vector search suits analytical corpora already in the warehouse. Choosing the store where the data already sits usually beats adding another system.
53. How do you choose a chunking strategy?
Fixed-size with overlap as a baseline (roughly 300–800 tokens, 10–15% overlap), improved by chunking on document structure so each chunk is coherent. Preserve metadata — title, section, page — for filtering and citation. Chunk size is the highest-leverage RAG parameter and must be tuned against an evaluation set, not guessed.
54. Why is hybrid search better than pure vector search?
Vector retrieval handles paraphrase and semantics; keyword retrieval reliably matches exact identifiers, part numbers, names and acronyms that embeddings blur. Fusing both catches queries that either alone would miss. Add reranking on top and you have the default production configuration.
55. What is the RAG Engine?
A managed RAG orchestration layer on the platform — corpora, file ingestion, chunking, embedding and retrieval exposed as an API, with pluggable vector backends. It sits between hand-building a pipeline and adopting Vertex AI Search wholesale.
56. How do you enforce security trimming in RAG?
Store permitted principals or groups as filterable metadata on each document and derive the filter from the caller's identity at query time. Never rely on the prompt to withhold content the retriever already returned — if it reached context, treat it as disclosed.
This separates candidates. Many answer "instruct the model not to reveal it." That is not access control, and interviewers are listening for exactly this.57. Your RAG returns irrelevant chunks. Diagnose it.
Test retrieval in isolation from generation first. Then check: chunk size too large (topic dilution) or too small (lost context); embedding model mismatch between index and query; no keyword component for identifier queries; no reranking; missing metadata filters; and query formulation — raw user questions often need rewriting before embedding.
58. How do you keep an index fresh?
Scheduled connectors with incremental sync, deletion propagation so removed sources leave the index, and a rebuild path for schema changes. State an RPO — "searchable within 15 minutes" is an answer; "we re-index nightly" invites the question of what happens at 9am.
59. When is RAG the wrong tool?
When the answer requires aggregation across the whole corpus — "how many contracts expire this quarter" is a SQL query, not a retrieval. When the task is behavioural rather than factual. And when the corpus fits comfortably in context, where retrieval only adds a failure mode.
Part 7 — Agents, ADK, A2A & MCP
60. What is the Agent Development Kit?
Google's code-first framework for building agents, now using a graph-based model organising agents into networks of sub-agents so multi-agent logic is defined explicitly rather than emerging. ADK v1.0 is stable across four languages, is open source, and deploys to Agent Engine or anywhere you can run a container.
61. Why does the shift to a graph-based framework matter?
Conversation-driven multi-agent systems are non-deterministic and hard to test — the same input can take different paths. A graph makes control flow explicit, so you can reason about it, test it, and debug a specific edge. It is the same reason workflow engines beat ad-hoc callbacks.
62. What is A2A, and how does it differ from MCP?
MCP standardises how a model reaches tools and data. A2A standardises how agents talk to each other — capability discovery, task delegation, status and artefact exchange across vendors and frameworks. They are complementary layers, not competitors, and being able to say that cleanly is a differentiator.
Current fact worth citing: A2A reached v1.0 and is reported in production at around 150 organisations.63. What is an Agent Card?
The A2A metadata document describing an agent — identity, capabilities, endpoint, authentication requirements — so other agents can discover and invoke it without prior integration. It is the mechanism that makes cross-vendor agent interoperability possible.
64. What are managed MCP servers, and what is Apigee's role?
Google provides managed MCP servers so tools are hosted and governed rather than self-run, with Apigee acting as an API-to-agent bridge — exposing an existing API estate as agent-consumable tools with the authentication, quota and observability Apigee already provides. For enterprises with hundreds of APIs, this is the shortest path to agent-ready infrastructure.
65. Single agent or multi-agent — how do you decide?
Start single. Split only for genuinely distinct responsibilities, different tool permissions, or different models suited to different sub-tasks. Multi-agent multiplies latency, cost and failure surface, and most systems described as multi-agent would work better as one agent with well-designed tools.
What impresses: arguing against multi-agent. Interviewers hear enthusiasm constantly; judgement is rare.66. What orchestration patterns does ADK support?
Sequential pipelines, parallel fan-out with aggregation, loop agents with exit conditions, and LLM-driven dynamic routing where a coordinator selects sub-agents. The graph framework makes these composable. Choose by the shape of the work, not by novelty.
67. How do you stop an agent taking a destructive action?
Least-privilege tool design with read and write as separate tools and scoped credentials per tool. Human approval for irreversible operations. Idempotency keys and dry-run modes. Step and spend budgets. Full tracing. Agent Identity so actions are attributable to the agent rather than a shared service account. And the governing rule: an agent's effective privilege must never exceed the trust level of the least-trusted content in its context.
68. What does production failure handling look like for an agent?
Tool-level timeouts and backoff, a maximum step budget to break loops, fallback to a simpler model or a canned response, structured errors returned to the model so it can recover rather than hallucinate, a circuit breaker on repeated failure, and alerting on step-budget exhaustion — usually the first signal something has regressed.
69. How do agents get exposed to end users?
Register an ADK agent hosted on Agent Engine, or an A2A-built agent, with Gemini Enterprise so it appears in the web app for employees. Alternatively expose it through your own application, Workspace via Workspace Studio, or a partner surface from the Agent Marketplace.
70. How do you evaluate an agent as opposed to a single response?
Evaluate the trajectory: did it resolve the user's intent, adhere to the task, choose the right tools, pass correct parameters, recover from failures? Final-answer-only evaluation hides the agent that got the right answer by luck after six wrong tool calls — which will break the moment inputs shift.
Part 8 — Tuning & optimisation
71. What tuning methods are available?
Supervised fine-tuning on prompt–response pairs (parameter-efficient, LoRA-style, rather than full-weight for Gemini), distillation from a larger model into a smaller one, and reinforcement learning from human feedback where preference data exists. Open Gemma weights can be fully fine-tuned since you control them.
72. How much data do you need?
Meaningful gains typically begin around 100 high-quality examples and improve into the low thousands, after which returns flatten. Consistency beats volume — a hundred examples that agree outperform a thousand that contradict each other.
73. What is distillation and when is it right?
Training a smaller model on a larger model's outputs for a narrow task. Right when you have a working expensive solution, stable requirements and high volume. Wrong while the task is still changing, because you will be re-distilling every sprint.
74. What is a prompt optimiser?
Tooling that automatically rewrites and tests prompt variants against your evaluation data, and helps migrate prompts across model versions. It is the practical answer to model churn — when a new Gemini version ships, you re-optimise rather than hand-tune.
75. Give me the optimisation ladder in cost order.
Better prompt → few-shot examples → structured output → context caching and prompt ordering → RAG → drop to a Flash tier with evaluation → routing → Batch for offline work → distillation → tuning → Provisioned Throughput. Work down the list, not up. Most teams start at tuning and spend ten times what they needed.
Part 9 — Evaluation & observability
76. What does the Gen AI Evaluation Service provide?
Two metric families. Computation-based — exact match, BLEU, ROUGE, tool-call accuracy — deterministic and cheap. Model-based — an autorater scoring coherence, fluency, groundedness, instruction-following, safety and question-answering quality against a rubric. Plus custom metrics you define.
77. What is an autorater and what are its limits?
A model scoring outputs against a rubric — scalable and reasonably correlated with human judgement. Limits: position bias, verbosity bias, self-preference for its own outputs, and run-to-run instability. Mitigate with a tight rubric, a different judge model than the one under test, and periodic human calibration on a sample.
78. How do you build a golden dataset?
Sample real queries across the distribution including edge cases and known failures. Have domain experts write or approve expected outputs. Keep it small enough to run per commit (100–300 items) with a larger nightly set. Version it. Add every production failure to it — that is how the suite stays honest.
79. What does observability mean for an AI application on GCP?
Distributed tracing across model calls, retrievals and tool invocations via Cloud Trace with OpenTelemetry, structured logs in Cloud Logging, token and latency metrics in Cloud Monitoring, plus Agent Observability for agent-specific traces, and continuous online evaluation sampling live traffic. Without traces you cannot answer "why did it say that," which is the only question anyone ever asks.
80. What is Model Monitoring?
Detection of training-serving skew and feature or prediction drift on deployed models, with alerting on threshold breach. Classical ML roles will probe this; for generative systems the equivalent is continuous evaluation on sampled traffic.
Part 10 — Responsible AI & security
81. What are Google's AI Principles, in practice?
Be socially beneficial, avoid creating or reinforcing unfair bias, be built and tested for safety, be accountable to people, incorporate privacy by design, uphold scientific excellence, and be made available for uses consistent with these principles. Interviewers ask to see whether you can connect a principle to a control you actually implemented.
82. What is SynthID?
Google's imperceptible watermarking embedded in generated images, audio, video and text, with a detection capability. It matters for provenance and for regulatory disclosure obligations, and it is a differentiator worth naming.
83. What is Model Armor?
A screening layer for prompts and responses covering prompt injection and jailbreak detection, sensitive data protection, malicious URL detection and content safety — applied consistently across models rather than relying on each model's own filters.
84. What is SAIF?
Google's Secure AI Framework — a structured approach covering supply chain, data poisoning, model theft, prompt injection and operational controls across the AI lifecycle. Naming it signals you think about AI security as a discipline rather than as a content filter.
85. How do you handle PII in an AI application on GCP?
Detect and de-identify with Sensitive Data Protection (formerly Cloud DLP) before the model sees it. Question whether the use case needs the raw value at all. Do not log raw prompts containing PII. Apply retention limits. And remember agent memory in Memory Bank is personal data subject to deletion requests.
86. Is customer data used to train Google's models?
No. Data submitted to the platform is not used to train Google's foundation models and is not shared with other customers, under the enterprise data governance commitments. Be precise here rather than approximate — it is a compliance question and imprecision is noticed.
87. How do you authenticate without service account keys?
Workload Identity Federation for external workloads, attached service accounts for GCP compute, and Application Default Credentials in code. Exported JSON keys are a finding in every security review — say you avoid them by default.
88. What does the EU AI Act mean for an Indian delivery team?
Risk-tiered obligations: prohibited uses; high-risk uses (employment, credit, education, biometrics) requiring risk management, data governance, logging, human oversight and conformity assessment; limited-risk requiring disclosure. It matters here because clients in scope push these obligations down the supply chain into delivery contracts.
Part 11 — Classical ML & MLOps
89. What are Pipelines and what do they run on?
Managed orchestration of ML workflows built with Kubeflow Pipelines or TFX, running as serverless DAGs with artefact lineage and caching. Caching matters — re-running a pipeline should not retrain an unchanged step.
90. What is Feature Store and what problem does it solve?
Centralised feature storage serving both training and online inference from the same definitions, which is how you eliminate training-serving skew. It also gives point-in-time correct retrieval so you do not leak future data into training labels.
91. What is training-serving skew?
Divergence between how features are computed in training versus serving — different code paths, different aggregation windows, different null handling. It is the most common cause of a model that scores well offline and fails in production. Feature Store exists largely to prevent it.
92. What is Model Registry?
Versioned model storage with metadata, lineage, aliases and deployment tracking — the control point for promoting a model through environments and for rolling back when a new version underperforms.
93. Online endpoint versus batch prediction?
Online for low-latency real-time serving with autoscaling and traffic splitting for blue-green or canary. Batch for large asynchronous scoring writing to Cloud Storage or BigQuery. Choose by latency requirement and payload size.
94. What is AutoML and when is it appropriate?
Automated model training on tabular, image, text or video data without writing training code. Appropriate for a strong baseline quickly, for teams without ML engineers, and as a benchmark against which to justify custom modelling. Inappropriate when you need control over architecture or inference cost.
95. How do you do explainability?
Vertex Explainable AI with feature attributions — sampled Shapley for tabular, integrated gradients and XRAI for images — surfaced at prediction time and in the model evaluation view. For regulated use cases in Indian BFSI this is frequently a hard requirement, not a nice-to-have.
Part 12 — Infrastructure, TPUs & cost
96. What is a TPU and when would you choose one over a GPU?
A Tensor Processing Unit — Google's custom accelerator optimised for large matrix operations. Choose TPUs for large-scale training and high-volume inference on models with good TPU support (JAX, TensorFlow, increasingly PyTorch/XLA), where price-performance at scale wins. Choose GPUs for broader framework and kernel compatibility, custom CUDA work, and smaller or more experimental workloads.
97. Name the TPU generations you know.
v5e (cost-efficient inference and mid-scale training), v5p (large-scale training), Trillium (v6e), and Ironwood (v7) — the inference-oriented generation. Knowing that the recent generations split along training versus inference optimisation is more useful than reciting FLOPS.
98. What is AI Hypercomputer?
Google's integrated supercomputing architecture — TPUs and GPUs, high-bandwidth interconnect, optimised storage, and a software stack spanning JAX, PyTorch/XLA and orchestration through GKE. It is the framing to use when asked about training at frontier scale.
99. How does the Agent Platform bill?
Pay-as-you-go across several dimensions rather than a flat subscription: Agent Engine runtime per vCPU-hour and per GB-hour of memory, session and memory storage per thousand events, Vertex AI Search per thousand queries by tier, and foundation model tokens priced separately by model. New accounts get $300 in credits for 90 days, and Express Mode allows trial without enabling billing.
Why this comes up: the runtime and storage dimensions surprise teams who budgeted only for tokens. Knowing all four lines is a genuine architect-level answer.100. Your GenAI costs are three times budget. What do you do?
Instrument first — cost per request by feature and model, or you are guessing. Then: check whether a Flash tier passes evaluation; enable context caching and reorder prompts so stable content sits first; trim retrieved context and system instruction; cap output tokens; move offline work to Batch; add routing; cache repeated queries. Only then consider Provisioned Throughput or distillation. Most overspend is retrieved context nobody measured.
101. How do you plan quota and capacity?
Estimate peak requests per second, average input and output tokens, derive tokens per minute, and compare against regional quota. Load test before launch. Plan a fallback region. Budget for the fact that real prompts are always longer than prototype prompts.
Part 13 — Scenario questions (3–8 years)
102. Build a support assistant over 60,000 internal documents with strict access control.
Ingest to Cloud Storage with document-level permission metadata. Use Vertex AI Search with the layout parser, hybrid retrieval and reranking. At query time derive an ACL filter from the caller's identity so retrieval is trimmed before generation. Ground responses with mandatory citations. Model Armor on input and retrieved content for injection. VPC Service Controls around the perimeter. Evaluate groundedness and retrieval quality against a golden set in CI. Trace to Cloud Trace. Serve through Agent Engine with per-team quota.
103. A stakeholder says the assistant "makes things up." Investigate.
Get specific failing examples — "hallucinating" usually turns out to be retrieval missing the document. Replay each through the trace: did retrieval return the right chunk? If no, it is a retrieval problem — chunking, hybrid search, reranking, parsing. If yes and the answer still diverged, tighten the instruction to require citations and to decline when sources are insufficient, drop temperature, and gate on groundedness. Then add every case to the golden set.
104. Design a multi-agent system for loan application processing.
Model it as an ADK graph, not a conversation: a document agent (Document AI extraction), a verification agent (external checks), a risk agent (a scoring model, not an LLM — regulated decisions need explainability), and a coordinator. Each sub-agent gets its own Agent Identity and least-privilege credentials. Human approval before any decision affecting the applicant, with full trace retention for audit. Explainable AI on the risk model. And be ready to argue that the credit decision itself should not be an LLM output.
What they want: that you know where not to put a language model. This question fails candidates who put Gemini in the decision path.105. Latency is 9 seconds; the business wants under 2. What changes?
Measure the span breakdown first — retrieval, model, tools. Then: stream so time-to-first-token is what the user feels; drop to a Flash tier; reduce thinking budget; cut retrieved context; parallelise independent retrievals and tool calls; enable context caching; use Provisioned Throughput if queueing is the cause; move non-essential work after the response. Also interrogate the requirement — for research-style tasks users tolerate latency when progress is visible.
106. The business wants to fine-tune Gemini on 300 PDFs so it "knows our products."
Push back with reasoning. Tuning teaches format and behaviour, not factual recall — and the products change next quarter, requiring retraining each time. RAG gives current answers with citations and updates by re-indexing. Propose RAG first, measure, and consider a light tune later only for tone or output shape. This question is testing whether you will tell a stakeholder they are wrong.
107. Your team has 40 agents built by different squads and nobody knows what exists.
This is exactly what the governance pillar addresses. Register everything in Agent Registry with named owners and versions. Give each agent an Agent Identity so actions are attributable rather than hidden behind shared service accounts. Route external access through Agent Gateway. Wire Agent Observability for traces. Then apply a standard: no agent reaches production without an owner, an evaluation set and a documented tool permission scope.
108. Migrate an existing Vertex AI workload to the Agent Platform.
Mostly nothing breaks — existing APIs are backward compatible and the console rebrands without manual migration. The real work is deciding what to adopt: enable the Agent Platform plan if you want Agent Runtime, Memory Bank, Agent Registry or the optimisation suite. Then migrate incrementally — register existing agents first for visibility, adopt Agent Identity next, and move runtime last. Do not rewrite working agents for the sake of the new brand.
109. Regulated Indian client, data cannot leave the country.
Deploy in asia-south1 or asia-south2 and verify model availability there — the newest models may not be present, so model choice is constrained. Regional not global endpoints. VPC Service Controls, CMEK, org policy restricting resource locations, Private Service Connect. Sensitive Data Protection before anything reaches a model. And set expectations early that they may run a slightly older model — better surfaced in week one than month six.
Part 14 — SDK & code round
110. Call Gemini from Python.
from google import genai
from google.genai import types
client = genai.Client(vertexai=True, project="my-project", location="asia-south1")
resp = client.models.generate_content(
model="gemini-2.5-flash",
contents="Summarise the Q3 revenue drivers from the attached report.",
config=types.GenerateContentConfig(
system_instruction="Answer only from the supplied context.",
temperature=0,
max_output_tokens=800,
),
)
print(resp.text)Pin the model version explicitly in production. Check resp.candidates[0].finish_reason before using resp.text — a safety block returns no text and naive code raises.
111. Force structured JSON output.
schema = {
"type": "object",
"properties": {
"sentiment": {"type": "string", "enum": ["positive", "neutral", "negative"]},
"themes": {"type": "array", "items": {"type": "string"}},
"confidence":{"type": "number"},
},
"required": ["sentiment", "themes", "confidence"],
}
resp = client.models.generate_content(
model="gemini-2.5-flash",
contents=review_text,
config=types.GenerateContentConfig(
response_mime_type="application/json",
response_schema=schema,
temperature=0,
),
)112. Define a tool and handle the call.
get_order = types.FunctionDeclaration(
name="get_order_status",
description="Return the current status of an order by ID.",
parameters={
"type": "object",
"properties": {"order_id": {"type": "string"}},
"required": ["order_id"],
},
)
resp = client.models.generate_content(
model="gemini-2.5-flash",
contents="Where is order A-4471?",
config=types.GenerateContentConfig(
tools=[types.Tool(function_declarations=[get_order])],
),
)
call = resp.candidates[0].content.parts[0].function_call
if call and call.name == "get_order_status":
result = lookup_order(call.args["order_id"]) # YOUR code executes, not the model113. Build a minimal ADK agent.
from google.adk.agents import Agent
def get_order_status(order_id: str) -> dict:
"""Return the current status of an order."""
return {"order_id": order_id, "status": "in_transit", "eta": "2026-08-18"}
root_agent = Agent(
name="orders_assistant",
model="gemini-2.5-flash",
instruction="You help staff check order status. Always use get_order_status.",
tools=[get_order_status],
)ADK reads the docstring and type hints to build the tool schema — so the docstring is production code, not a comment.
114. Handle rate limits correctly.
import time, random
from google.api_core import exceptions
def call_with_retry(fn, max_attempts=5):
for attempt in range(max_attempts):
try:
return fn()
except exceptions.ResourceExhausted:
if attempt == max_attempts - 1:
raise
time.sleep((2 ** attempt) + random.uniform(0, 1)) # backoff + jitterInterviewers look for exponential growth, jitter, and a bounded attempt count. Then say what you would change structurally — retry alone is not an architecture.
115. Run an evaluation.
from vertexai.evaluation import EvalTask, MetricPromptTemplateExamples
task = EvalTask(
dataset=eval_df,
metrics=[
MetricPromptTemplateExamples.Pointwise.GROUNDEDNESS,
MetricPromptTemplateExamples.Pointwise.INSTRUCTION_FOLLOWING,
"rouge_l_sum",
],
experiment="support-assistant-eval",
)
result = task.evaluate()
print(result.summary_metrics)Wire the thresholds into Cloud Build so a regression fails the pipeline rather than reaching users.
Part 15 — Salary bands and the HR round
Indicative 2026 ranges for Google Cloud AI / GenAI engineering roles in India. Product companies, GCCs and US-shift roles pay above these; service-company bands sit at the lower end.
| Experience | Typical title | Range (₹ LPA) |
|---|---|---|
| 0–2 years | AI/ML Engineer (Associate) | 5 – 10 |
| 2–4 years | GCP AI Engineer | 10 – 19 |
| 4–7 years | Senior AI / GenAI Engineer | 19 – 35 |
| 7–10 years | AI Architect / Lead | 35 – 58 |
| 10+ years | Principal AI Architect / Practice Head | 58 – 95+ |
Questions you should ask them
- Is anything live with real users, or is this still pilots?
- Have you moved to the Agent Platform plan, or are you still on the pre-rebrand feature set?
- Do you have an evaluation suite, and does it gate deployment?
- Who owns cost, and is spend measured per feature?
- Do agents have write access to production systems, and what is the approval model?
Certifications that carry weight in 2026
- Professional Machine Learning Engineer — the core certification for these roles
- Generative AI Leader — useful for consulting and pre-sales-facing roles
- Professional Cloud Architect — where the role spans platform design
- Professional Data Engineer — where the role touches BigQuery and pipelines
- Associate Cloud Engineer — a reasonable entry point for career changers
Prepare for these interviews with Cloud Soft Solutions
Our APEX program covers AI, GenAI, Cloud and Cyber Security — including a Google Cloud AI track spanning the Gemini Enterprise Agent Platform, ADK, A2A, RAG and Responsible AI, with live lab environments, capstone projects, mock interviews and resume preparation. 5,500+ alumni placed.
📍 513, 5th Floor, Aditya Enclave, Nilagiri Block, Beside Ameerpet Metro Station, Ameerpet, Hyderabad – 500016
📞 +91 96660 19191 · +91 99496 16388 · ✉️ info@cloudsoftsol.com
How to prepare (do not memorise this list)
- Build one RAG application end to end on your own documents, and measure it.
- Build one ADK agent with two tools, then deliberately break it — a failing tool, injected instructions in a retrieved document, an infinite loop. Fix each.
- Create a 50-item golden dataset and run the evaluation service against it.
- Do a cost teardown. Know your cost per request and where it goes across all four billing dimensions.
- Read the Google Cloud AI release notes monthly.
- Prepare three stories: something you shipped, something that failed in production and what you changed, and a time you talked a stakeholder out of the wrong approach.
Note on accuracy: Google Cloud's AI platform changes rapidly and the brand changed as recently as April 2026. Details here reflect the platform as of August 2026. Verify against Google Cloud documentation and release notes before any interview — and if an interviewer still says "Vertex AI," follow their vocabulary while showing you know the current name.