Google AI, ML, RAG and Agentic-AI Interview Questions (2026)

Google does not publish its interview questions. Every source behind this list is a prep vendor or an anonymous review site (Glassdoor, Blind, LeetCode Discuss). Use the topics as the signal and treat the exact wording as approximate.

How to read the tags

• [R] Reported: a candidate post or review, or a coaching vendor citing such reports.
• [V] Vendor-tagged: a prep site labels it as Google; the underlying report is not visible.
• [P] Prep-only: not Google-specific, or our own synthesis. Useful for breadth.
• [~] Practice: a practice scenario we could not trace to any source page.

The short version: the classic ML system-design questions are credibly reported by Google candidates. Most RAG, agent, and LLM-serving prompts online are good practice scenarios, not confirmed transcripts. Coding and a resume-driven ML deep dive carry as much weight as system design.

Verified ML system-design questions

• Build, train, and deploy a system that detects whether multimedia or advertising content violates the terms of service or is offensive. [R]
• Design autocomplete and/or spell-check for a mobile device. [R]
• Design autocomplete and/or automatic replies for email. [R]
• Design the YouTube recommendation system. [R]

Follow-up interviewers add

• How would you optimize prediction throughput for an RNN-based model? [R]
• Which loss function would you optimize, and why? [R]
• What data would you collect, and why? [R]
• How do you avoid bias and feedback loops? [R]
• How do you handle a corrupt model or an incorrect training batch? [R]

Further reported prompts

• Design a job-recommendation system. [R]
• Design type-ahead search. [R]
• Explain precision and recall. [R]
• Explain root mean square error. [R]

Additional Google-tagged ML questions

ML system design

• Design a content-moderation system. [R]
• Design a system that processes traffic data. [R]
• Detect ad boxes in YouTube videos, then code the frame ranges and positions. [R]
• Predict which links to show from user, timestamp, query and click history. [R]
• Design a real-time recommendation system for a consumer product. [V]
• Design a product or video recommender. [V]
• Design a feedback-driven recommender (contextual bandit, online learning). [V]
• Design large-scale near-duplicate video detection. [V]
• Design obstacle detection for a self-driving car. [V]
• Predict a reaction factor from molecule pairs (tabular, end-to-end). [V]
• Split a matrix multiplication across two GPUs. [V]
• Design an ML experiment tracking and analysis platform. [V]
• Design Gmail spam detection; Search autocomplete; a petabyte-scale training pipeline. [P]

ML depth and LLM fundamentals

• Walk through the full LLM lifecycle: data filtering, pretraining objectives, architecture, alignment. [V]
• Explain the Transformer and its variants in depth. [V]
• Logistic regression, Naive Bayes, multi-class metrics, bagging vs boosting. [V]
• Compare regularisation methods and when to use each. [V]
• Compare tokenization approaches; design LLM-based recommendation parts. [V]
• Explain GRPO-style reinforcement learning for diffusion models. [V]
• Implement a multi-head attention class, then discuss current architectures. [R]
• Sample efficiently from a multinomial distribution. [R]
• Implement k-means or logistic regression from scratch; convolution and batch norm in pure Python. [R]
• Precision and recall, over- and underfitting, regularisation, dropout. [R]
• Handling overfitting, when to stop training, and deploying on GCP or AWS. [R]
• Walk through a modeling challenge: detection, alternatives, fix, evidence. [V]
• Why might a model that is good offline fail online? [P]
• Your A/B test won, but your manager fears long-term harm. [V]

RAG questions

• Design a RAG question-answering bot: chunking, embeddings, vector-store choice, retrieval, response synthesis. [V]
• Design an enterprise RAG pipeline over millions of PDFs with retrieval under 50 ms and no context overflow. [V]
• Evaluate retrieval precision and recall separately from generation quality. [V]
• Compare dense and sparse retrieval; build and tune a hybrid system for low latency. [V]
• Detect, quantify and mitigate hallucination and prompt injection in a multi-step workflow. [V]
• Design an enterprise RAG pipeline on vector databases for unstructured data. [V] [~]
• Design RAG over confidential documents with strict per-user access control. [V] [~]
• Debug a cascading failure where an upstream agent passes malformed context downstream. [V] [~]
• Build an automated evaluation pipeline for regression, bias and hallucination in a summarisation tool. [V]
• Build a real-time validator that blocks hallucinated or out-of-schema output. [V] [~]
• Plan a RAG project and discuss the model lifecycle. [R]

Agentic-AI questions

• Design an agent that searches and synthesises unstructured internal documents with RAG. [V]
• Design an internal-search agent with accurate access control, no hallucination and real-time index updates. [V]
• Architect a story-writing agent with long-range consistency, character memory and state tracking. [V]
• Design a BigQuery agent that turns natural language into optimised multi-stage pipelines. [V]
• Define tools, function calling and recovery for unexpected API failures. [V]
• Architect an agentic cloud-security system that detects, analyses and mitigates incidents in real time. [V]
• Design a multimodal smart-glasses agent that plans actions from camera and speech. [V]
• Handle tool-calling failures and non-deterministic loops in production. [V]
• Design a multi-agent system that analyses security logs, isolates compromised assets and writes incident reports. [V]
• Build an evaluation suite for trajectory accuracy, tool-selection precision and hallucination rate. [V]
• Build a benchmark checking whether an analytics agent chose the right SQL tools and schema parameters. [V]
• Task success drops after a foundation-model checkpoint update. What do you do? [V]
• Design infrastructure for an enterprise agent serving millions of concurrent requests. [V]
• Manage and invalidate context caches over dynamic documents. [V]
• Implement a thread-safe conversation buffer with token truncation. [V]
• Given tool dependencies, detect cycles and output a valid order. [V]

LLM serving and ML infrastructure

• Design a distributed LLM serving engine on TPU/GPU clusters: KV-cache compression, PagedAttention, continuous batching, multi-region balancing. [V]
• Serve an ads LLM at 100,000 QPS with p99 under 100 ms. [V]
• Design a Spotify- or Google-Play-style real-time recommender across candidate generation, ranking and re-ranking. [V]
• Combine real-time feature retrieval with batch deep-learning inference in a music recommender. [V]
• Add an LLM agent for natural-language follow-ups to a music recommender. [V]
• Design a multi-stage recommender with production monitoring. [V]
• Cut latency and compute when deploying multimodal generative models to edge devices. [V]
• Handle cold start and real-time context in large recommenders. [V]
• Route queries between lightweight and frontier models under cost budgets. [V] [~]
• Reduce latency and token cost without hurting quality. [V] [~]
• Serve multimodal Gemini models across distributed edge nodes. [V] [~]

ML theory and evaluation

• Derive cross-entropy and hinge loss, with gradients and loss-landscape properties. [V]
• Design an evaluation framework for a production generative app; contrast ROUGE/BLEU with LLM-as-judge and human review. [V]
• Compare LoRA and QLoRA with full fine-tuning: theory and memory savings. [V]
• Trade-offs among cross-entropy, triplet loss, and DPO for tool-selection fine-tuning. [V]
• Explain attention and how context compression affects retrieval accuracy. [V]
• Which metrics monitor drift and degradation in production agents? [V]
• Design continuous evaluation for trajectory accuracy, tool selection, and hallucination. [V]
• How do loss functions differ for free text versus structured tool calls? [V]

Coding questions

Agent- and retrieval-flavored problems

• Given N CPUs and M task durations, find the minimum completion time. [V]
• Count valid unlock patterns of length at least L on an N x M grid with eight-direction moves and no jumping over unvisited dots. [V]
• Find the most frequent event and its timestamps; extend to a sliding window. [V]
• Build a balanced search tree or graph from a sorted array under structural constraints. [V]
• Topologically sort subtasks; traverse an N-ary tree in place. [V]
• Weighted directed graph: can the source reach the target, and what is the shortest path? [V]
• Implement a QuadTree with insert, point, and range queries. [V]
• DP: choose the action sequence maximizing cumulative reward under state constraints. [V]
• Thread-safe conversation buffer with token truncation. [V]
• Detect circular tool dependencies and return a valid order. [V]
• Embedding cache to avoid redundant vector searches. [V] [~]
• Parse streaming LLM output and enforce a JSON schema. [V] [~]
• Dynamic chunking and token distribution for long documents. [V] [~]
• Thread-safe state manager for agent execution history. [V] [~]

Additional reported problems

L6 report

• Logger or timer with added constraints
• Grid problem from O(n^2) to O(n)

Vendor-tagged

• Substring search with weighted sampling
• Streaming weighted sampler
• Connected crop layout and safe paths
• Fence-painting DP
• Robust Python generator
• LRU-cache variant
• Trie autocomplete
• Multi-tenant rate limiter

Reported

• Maximum path sum in a binary tree
• Decode string
• Word ladder
• Longest increasing path in a matrix
• Snapshot array
• Minimum window substring
• Maximal rectangle

Interview Query list

• String shift
• Bigrams
• Bracket matching
• 2D rain water
• Nearest common ancestor
• Minimum change
• Basic regex (. and *)
• First to roll a six
• 500 cards, three draws in increasing order
• Explain a p-value to a layperson

Product, behavioral, and responsible AI

AI-specific

• Choose MVP functionality for a complex enterprise GenAI case. [V] [~]
• How do you decide a prototype is ready for production? [V] [~]
• Evaluate cost per request, time to first token and tokens per second against UX SLAs. [V] [~]
• A feature has high engagement but unacceptable hallucination or safety risk. What now? [V] [~]
• Turn a fragile GenAI prototype into a fault-tolerant product. [V] [~]
• Balance accuracy against latency under a tight deadline. [V]
• Resolve conflict between research recommendations and infrastructure limits. [V]
• Describe a deployed AI system that failed; diagnosis and prevention. [V]
• A collaborator took public credit for your work. [V]
• You had several P0 tasks. How did you realign stakeholders? [V]

Googleyness and general

• Conflict, failure and proudest-project stories using STAR. [V]
• What hurdles have you hit in data projects? [R]
• How do you handle conflict and ambiguity? [V]
• Present and defend your research and explain your interests (intern). [V]
• Why Google? Favourite Google product? Conflict on your team? [R]
• How do you approach GenAI safety in consumer products? [P]

---

Cross-company LLM, RAG, and agent questions

• Single-GPU batching for up to 100 synchronous inputs. [P]
• End-to-end LLM batching; efficient sampling API. [P]
• Insurance-claims agent with RAG and token-cost control. [P]
• Customer-service agents (outdoors retailer; streaming service). [P]
• The model is confidently wrong in high-risk settings. [P]
• Explain RAG; prompting vs RAG vs fine-tuning. [P]
• LLM-as-judge failure modes; catching regressions on a model swap. [P]
• Async retries with timeouts and exponential backoff. [P]
• Detect circular references among spreadsheet cells. [P]
• In-memory key-value store with begin, commit and abort. [P]
• RAG-based agent; agent harness and evaluation system. [P]
• Distribute model weights to a GPU fleet. [P]

Practice drills (our synthesis, no interview evidence)

• Estimate GPU memory to serve a 70B-parameter model at FP16 versus INT8 with 32 concurrent 8K-token requests; separate weights from KV cache. [P]
• When does speculative decoding help, and when does it hurt? [P]
• Compare post-training quantisation with quantisation-aware training; which accuracy risks do you test? [P]
• Implement reciprocal rank fusion for hybrid retrieval and explain tuning. [P]
• How can a retrieved document carry a prompt injection? Defences at each stage? [P]
• When would you use MCP versus A2A? [P]
• Design an orchestrator with sub-agents; mark idempotent calls, approval gates and rollback. [P]
• Design invalidation for a semantic cache over live documents. [P]
• Validate an LLM judge against human labels before using it in CI. [P]
• Triage a task-success drop after a checkpoint update. [P]

Full guide: Link