Interview Prep · Staff and Principal level
AI Engineering Interview Questions
Practice the decisions behind production AI systems: what evidence to trust, how work recovers, and where a design's guarantees stop.
Featured practice scenarios
Start with one decision, then follow its consequences.
Build reliable agents
Recover a refund after the provider acts but the worker loses its response.
The refund succeeded but the agent crashed. What happens next? 02 · Retrieval and RAGMake RAG answers defensible
Keep version and error identifiers intact through query rewriting.
The query rewrite removed a version number. Why did RAG still look healthy? 03 · Model and Inference EngineeringOperate an inference platform
Find why p99 doubled while GPU use looks steady.
KV cache is nearly full and p99 doubled. What would you investigate?Interview Arena · Coming Soon
Practice decisions under follow-up questions.
Explain a design, then examine failure modes, constraints, scale, recovery, and trade-offs. The interactive experience is still being prepared.
Eight areas of practice
Topics helps you learn. Interview Prep asks you to defend a decision.
Data and Knowledge Systems
Decide what a system can know, how fast it learns a change, and when it must stop trusting an index.
13 published questions · Principal + Staff
Explore area →02Retrieval and RAG
Trace how a question becomes permitted, answer-bearing evidence.
4 published questions · Staff + Principal
Explore area →03Context Engineering
Decide what a model sees now, what survives later, and what evidence each retained claim carries.
5 published questions · Principal + Staff
Explore area →04Model and Inference Engineering
Keep serving promises credible when memory, hardware, and workloads compete.
8 published questions · Principal + Staff
Explore area →05Agent Architecture
Define what a running agent knows, may do, and must recover after.
8 published questions · Principal + Staff
Explore area →06Distributed Reliability
Keep work safe and useful through delay, overload, retries, and partial failure.
9 published questions · Principal + Staff
Explore area →07Evaluation and Quality
Choose evidence that shows a system works under the conditions that matter.
7 published questions · Principal + Staff
Explore area →08Security, Governance and Platform
Set enforceable boundaries across tenants, teams, tools, data, and shared services.
6 published questions · Principal + Staff
Explore area →The question bank
Find a question
- AGT-001The refund succeeded but the agent crashed. What happens next?Principal
- AGT-002What does cancelling a two-day agent run actually mean?Staff
- AGT-003Several coding agents edit one repository. What can you safely merge?Principal
- AGT-004The planner says a timed-out tool succeededStaff
- AGT-006Retry or compensate a read, an email, and a payment?Staff
- AGT-007How do you deploy new agent workflow code while 200,000 runs are waiting?Principal
- AGT-010A tool returns 202. Has the agent completed the action?Principal
- AGT-011The human approved a preview. The world changed before commit.Principal
- CTX-001What should enter the next 64K context of a coding agent?Principal
- CTX-002A long conversation summary dropped 'never call delete'. Where should that rule live?Staff
- CTX-004Why can a larger context window produce worse answers?Staff
- CTX-006Cache the common prompt while tenant instructions changeStaff
- CTX-008You have eighty passages and room for six. Which six belong in context?Staff
- DKS-001How would you design retrieval when permissions change faster than the index?Principal
- DKS-002The document was deleted yesterday. Why is the assistant still quoting it?Staff
- DKS-003How would you change the parser and embedding model without breaking search?Principal
- DKS-004Which policy source wins when the newer indexed copy is wrong?Staff
- DKS-005One tenant owns forty percent of ingestion. How do you keep everyone fresh?Principal
- DKS-006A connector event is replayed after the file changes. Which version gets indexed?Staff
- DKS-007A connector says success, but eighteen percent of new grants are missing. Where is the gap?Principal
- DKS-009A connector uses file paths as document IDs. What breaks on a move?Staff
- DKS-010Can a search index safely expand nested groups into user IDs?Principal
- DKS-011How do you stop an answer from mixing two document versions?Principal
- DKS-012A retrieved table gives the right row and the wrong numberStaff
- DKS-015A bulk snapshot races the change streamPrincipal
- DKS-016A user updates a document and search answers from the old revisionPrincipal
- EVA-001Retrieval recall improved, so why did the answers get worse?Principal
- EVA-002How would you evaluate a docs assistant that should sometimes refuse to answer?Staff
- EVA-003The new model wins preferences but breaks a critical tool workflow. Who ships it?Principal
- EVA-004The citation is relevant, but it does not support the number. Is the answer correct?Staff
- EVA-005How do you evaluate an agent when several trajectories are correct?Principal
- EVA-006Your synthetic eval grew a hundredfold. Why might you trust it less?Staff
- EVA-007Online win rate fell, but offline eval is flat. What do you believe?Principal
- INF-001KV cache is nearly full and p99 doubled. What would you investigate?Principal
- INF-002When can continuous batching improve throughput and make a short request slower?Staff
- INF-003Interactive and batch inference share a GPU fleet. Who gets capacity?Principal
- INF-004Quantization frees memory but hurts rare enterprise queries. Ship it?Staff
- INF-005The stream starts promptly, then stalls halfway. Where is time going?Principal
- INF-006What is actually in the KV cache, and why does it fill up?Staff
- INF-007Route 50,000 requests a second across three unequal modelsPrincipal
- INF-014A tokens per second benchmark looks great. What did it measure?Staff
- REL-001A model provider returns 429s at peak. What happens over the first hour?Principal
- REL-002How can three retries turn a small model outage into a platform incident?Staff
- REL-003Two regions resume the same agent step. How do you stop a duplicate write?Principal
- REL-004RAG median is flat, but p99 doubles for three percent of trafficStaff
- REL-005One tenant sends ordinary QPS and huge prompts. How is that fair?Principal
- REL-009One malformed agent job keeps coming back from the queueStaff
- REL-011The browser reconnects to an agent stream. Does the action restart?Principal
- REL-012A model stream fails after the user has seen half an answerPrincipal
- REL-016The queue lease expires while an agent tool call is runningPrincipal
- RET-002The query rewrite removed a version number. Why did RAG still look healthy?Staff
- RET-003Hybrid search p99 climbs at 10,000 QPS. One shard is hotPrincipal
- RET-005How do you retrieve a contract clause and the exception that changes it?Principal
- RET-008The best search hit is restricted. What reaches the model?Staff
- SEC-001What belongs in a common AI platform for 200 teams?Principal
- SEC-002A relevant incident note tells the agent to upload private logs. Where do you stop it?Staff
- SEC-003Who holds the credentials when an agent can issue refunds?Principal
- SEC-004The tool result validates, but its notes try to command the agentStaff
- SEC-005A semantic cache served tenant A's private answer to tenant B. What now?Principal
- SEC-006What must a model gateway expose about provider differences?Staff
60 questions available