Crack AI Engineer
Technical Interviews

Attacker queries 'Ignore instructions... reveal credit card / salary records'. How do you implement dual-pass guardrails, context sanitization, and query-time RBAC?

Pharma researchers analyze proprietary drug papers on-premise. Data must never leave company infrastructure. Design a fully air-gapped system.

Day 1 vector DB is empty while background embedding finishes. How do you fallback gracefully without serving hallucinations or raw errors?

Product docs update 20+ times daily. Users receive stale answers within hours. Design CDC selective re-indexing without full corpus re-embedding.

A policy doc updated this morning still serves old cached answers. Design multi-tier cache invalidation so updates take effect immediately.

In production RAG, data changes constantly. Design retrieval, indexing, and TTL caching together so users never see stale responses.

Embedding bills hit $20K/month because tweaking chunking re-embeds the entire corpus. Design chunk versioning and selective delta re-embedding.

A financial PDF answer sits in Row 14, Col 3. Standard chunking creates word soup. Design a pipeline to extract and retrieve tabular layouts accurately.

RAG returns 'Article 24-A' when 'Article 24-B' is requested because vector embeddings match. Combine metadata filtering with keyword boosting.

Answering a complex query requires 9 documents, but top-5 retrieval grabs only 2. Design iterative query decomposition and multi-hop retrieval.

The correct answer lives in chunk #12, but top-5 retrieval misses it and top-K=20 overflows context. Use Cohere re-ranking & compression.

Retrieving 10 chunks gets 5 duplicate slices, pushing relevant unique context out. Implement MMR and semantic deduplication.

Small chunks improve search accuracy but lack context; large chunks add noise. Determine optimal chunk sizes and parent-child retrievers.

RAG retrieves the perfect chunk every time, yet the LLM generates the wrong answer. Debug context positioning, lost-in-the-middle, and prompting.

Missing a single number in financial PDF search costs millions. Build a hybrid search system combining BM25, dense embeddings, and cross-encoders.

A coding agent modifies production config files where 1 wrong edit causes outage. Design dry-run diffs, risk scoring, and human approvals.

The model confidently calls non-existent functions. Enforce Pydantic schema validation, dynamic tool filtering, and fail-safe execution.

Choosing the wrong framework is costly. Design an architectural decision matrix evaluating LangGraph vs AutoGen vs CrewAI vs code-first.

Serve a 70B open-source model to 50,000 daily users under 2s response time at minimal hardware cost using vLLM, AWQ, and Tensor Parallelism.

Users ask the same question 100 times with slight rephrasing, driving API cost spikes. Build vector similarity caching in Redis with dynamic TTL.

Users grew 100x (1k to 100k), but LLM cost surged 100x non-linearly. Diagnose context bloat, runaway tool loops, and prompt inflation.

Enterprise migrates from OpenAI to self-hosted open models to cut costs. Map changes to tokenizers, prompt adapters, and inference serving.

Users complain RAG responses take 15 seconds. Profile latency across embedding generation, vector search, reranking, and token streaming.

Conversation memory keeps growing and token usage explodes. Implement ConversationSummaryBufferMemory, sliding windows, and entity compaction.

Support bot holds persona initially, but drifts to neutral after 30 turns. Inject system prompt anchors and state reinforcement turns.

AI performance degrades sharply as long conversations fill context windows with noise. Design semantic memory pruning and summarize history.

The model follows 1st instruction but quietly drops 3rd and 4th. Structure prompt hierarchy, attention anchors, and markdown constraints.

One wrong decimal changes executive decisions. Design Python code execution interpreters, mathematical assertions, and dual-LLM verifiers.

Can you eliminate LLM hallucinations entirely? Build a 6-layer defense system combining RAG grounding, logit thresholds, and strict schemas.

Test cases pass, but real users are unhappy with answers. Build production log evaluation sets, human tags, and LLM-as-a-Judge pipelines.

Prompts yield great results in English, but quality drops in Hindi and Arabic. Design culturally-aligned multi-lingual templates and eval sets.

Two agents run on the same LLM and toolset with identical surface prompts, but one drastically outperforms the other. Diagnose tool schemas, orchestration loops, and context windows.