Fine-Tuning Interview Questions for Experienced

**Training** Refusal datasets, toxicity filters on training data.

Model question on RLHF — SFT, reward model, PPO alignment, and how human preferences change helpfulness vs hallucination trade-offs.

Part 3 of the hallucination handbook — production defense layers from retrieval and prompting through guardrails, citations, confidence scoring, eval frameworks, and enterprise architecture. Asked at OpenAI, Google, Meta, Anthropic, Microsoft, and Amazon.

Model-based interview on DPO vs RLHF — preference pairs, training simplicity, and production trade-offs.

**Training** Rank selection, learning rate, epochs; validation loss + task evals.

**3.1** 128k context, stronger instruction following — update inference stack for long context KV cache.

**Use cases** Where zero-shot vision fails consistently on proprietary visual patterns.

**Decision** Try prompt + tools + RAG first; fine-tune if style/reliability gap persists and data exists.

Project question on synthetic data — distillation datasets, diversity, contamination, quality filters, and human review.

Company/scenario question on open vs closed LLMs — self-hosting economics, compliance, upgrade velocity, and hybrid strategies.

Hard model question comparing CAI, RLHF (PPO), and DPO — data efficiency, stability, and production fine-tuning trade-offs.

Hard scenario on LLM cost control — token budgeting, routing, caching, batch APIs, and unit economics at millions of queries.

Scenario question comparing prompting, RAG, and fine-tuning for domain knowledge — freshness, cost, auditability, and when to combine approaches.

Hallucination in RAG systems is the #1 production failure mode cited in AI engineering interviews. Your interviewer wants a systematic debugging framework — not a list of buzzwords. Learn how to measure faithfulness, fix retrieval precision, and layer mitigations the way senior engineers at Databricks and Meta actually ship RAG.

Enterprise RAG interviews test system design at scale: ACL-aware retrieval, audit logging, and ingestion pipelines for millions of documents. This is a staff-level question appearing at Microsoft, Salesforce, and Fortune 500 AI teams. Walk through a complete architecture with security boundaries and operational concerns.

Multi-agent orchestration is a senior/staff-level system design question gaining traction at Google DeepMind and Microsoft. Learn the supervisor pattern, shared state management, and how to avoid the 'too many agents' anti-pattern that sinks most candidate answers.

LangGraph has emerged as the go-to framework for stateful agent workflows, replacing brittle LangChain chains in production. Interviewers want graph thinking — nodes, edges, conditional routing, and checkpointing — not just API familiarity. Essential for any LangChain shop hiring GenAI engineers.

Structured JSON output from LLMs is a production necessity at Stripe, Vercel, and every AI-native startup. Interviewers test your reliability engineering — schema enforcement, retry logic, streaming edge cases — not just 'use JSON mode.' Master the full production playbook.

OpenAI's model lineup changes fast. GPT-4 vs GPT-4o is a model selection question that tests whether you understand latency, cost, multimodal capabilities, and when reasoning depth matters. Critical for any role touching OpenAI APIs in production.

Constitutional AI is Anthropic's differentiator and a must-know for Claude-focused interviews. Go beyond the marketing — explain the self-critique training loop, how CAI compares to RLHF, and practical safety implications for production deployments.

Google's Gemini 1.5 Pro long-context window opens use cases impossible with standard LLMs — whole-codebase analysis, multi-hour video, massive document review. Interviewers test whether you understand real limitations behind the 1M token marketing number.

Self-hosting Llama 3 is a infrastructure-heavy question for ML platform and AI engineer roles at Meta-adjacent companies. Expect deep dives on quantization, vLLM, GPU sizing, and the TCO math that determines build vs buy decisions.

500M vectors at sub-100ms p99 is a staff-level vector search design question from Uber, Airbnb, and large-scale ML platform teams. Learn sharding strategies, index tuning, and the operational trade-offs that separate senior from principal engineers.

Claude Code is Anthropic's CLI coding agent designed for repository-level work: reading the codebase, editing multiple files, running tests/builds, and iterating until a task succeeds. Unlike autocomplete-first copilots that live inside the editor, Claude Code operates as an agent loop — observe → plan → act (edit/run tools) → verify.