The AI Interview Process (What to Expect)

The standard 5-stage loop

This is the modal interview process at most AI-focused companies in 2026 (mid-sized startups to FAANG). Frontier labs add a research round; small startups often compress stages.

Stage 1: Recruiter screen (30 min)

What it tests: fit on basics — background, comp expectations, timeline, role match.

What to prepare:

• 2-minute "tell me about yourself" focused on engineering + AI work
• Why interested in their company specifically (research them; don't be generic)
• Salary expectations (have a number; don't say "negotiable")
• Timeline (when can you start)

Common mistakes: being vague on what kind of role you want; over-explaining technical depth (save for later stages); not asking about the team / interview process.

Stage 2: Technical phone screen (60 min)

What it tests: foundational technical depth + can-you-think-on-your-feet.

Format varies by company:

• Coding problem in a shared editor (45 min + 15 min discussion)
• AI concept discussion (RAG, fine-tuning, evals — pick one and go deep)
• Hybrid: 30 min concepts + 30 min coding

What to prepare:

• Practice coding in a shared editor (CoderPad, HackerRank) — different from your IDE
• Be ready to discuss tradeoffs in an AI concept ("when would you use RAG vs fine-tuning?")
• Talk through your thinking while you code; don't go silent

Common mistakes: solving silently; jumping to code before clarifying requirements; not handling edge cases.

Stage 3: Coding round (60-90 min)

What it tests: practical coding ability on a meaningful problem.

Typical problems:

• Implement scaled dot-product attention in NumPy
• Build a small RAG from scratch (retrieve + format + generate)
• Debug a broken training loop
• Implement a simple agent loop with one or two tools
• Practical algorithm (less common at AI-focused shops): graph traversal, dynamic programming

What to prepare:

• Implement attention from scratch (good signal you understand transformers)
• Build a small RAG in 30 minutes from a blank file (shows you can ship)
• Practice debugging — interviewers may give you broken code

Common mistakes: over-engineering; not asking about constraints; insisting on perfect solutions instead of iterating.

Stage 4: AI system design (60-90 min)

What it tests: judgment, architectural thinking, awareness of tradeoffs at scale.

Typical prompts:

• Design a customer support RAG chatbot
• Design a real-time fraud detection system
• Design an AI agent that books meetings
• Design a recommendation engine for a video platform
• Design an internal LLM evals platform

Structure your answer:

1. Clarify (ask 3-5 questions): scale, latency, what we're optimizing
2. Sketch the architecture (whiteboard or shared doc): high-level components
3. Drill into specifics: model choice, data flow, scale numbers, cost
4. Discuss failure modes: what breaks, how you'd monitor, how you'd debug
5. Talk tradeoffs: what you're not doing, why

What to prepare:

• Practice 5-10 system design prompts out loud (see /interviews System Design questions)
• Know the typical cost/latency numbers for LLM APIs
• Understand RAG architecture deeply (most prompts touch it)
• Know when to use classical ML vs LLMs

Common mistakes: jumping to model choice before clarifying; ignoring cost; over-engineering for scale you don't need; not mentioning evals.

Stage 5: Onsite (4-5 interviews × 45-60 min)

Typical composition:

• 1-2 coding rounds (similar to phone screen but harder)
• 1-2 system design rounds (deeper than the earlier one)
• 1 behavioral / collaboration interview
• 1 "executive" interview (with a director or VP) — often unstructured

Behavioral interview prep:

• 5-7 stories that show: technical impact, disagreement resolution, dealing with failure, ethics judgment, cross-functional work
• STAR format but tightened
• AI-specific questions: ethics dilemmas, handling hallucinations, agent safety decisions

Executive interview prep:

• Often "tell me about yourself" + open conversation
• They're testing: communication, judgment, what you care about
• Have a clear thesis on AI trends + your career goals
• Ask substantive questions about company direction

Frontier-lab additions

Companies like OpenAI, Anthropic, DeepMind, Mistral add:

Research deep-dive (60-120 min)

Pick a paper you've read and discuss it deeply:

• Summarize the contribution
• Walk through the experimental setup
• Critique the methodology
• Propose extensions / what you'd try next

What to prepare:

• Have 3-5 papers you can discuss at this depth (not just skim)
• Recent + classic: "Attention Is All You Need," recent flagship paper from the company you're interviewing at, a paper that influenced your project work

ML internals coding

Often: implement an attention layer from scratch in NumPy, train a tiny transformer on toy data, debug a gradient issue.

What to prepare:

• Karpathy's "Let's build GPT" video — implement along with it
• Read the original transformer paper code
• Comfort with PyTorch autograd internals

Timeline + offer expectations

4-8 weeks from first conversation to offer is typical:

• Week 1: recruiter screen
• Week 2: tech phone screen
• Week 3-4: coding + system design rounds
• Week 5-6: onsite
• Week 7-8: offer + negotiation

Smaller startups can move in 1-2 weeks; frontier labs and senior research roles can take 8-12 weeks.

Offer process:

• Verbal offer usually first
• Written offer with full comp 1-3 days later
• Negotiate via email — never accept on the phone
• Get competing offers if you can; they'll usually match or come close
• Expect 1-2 weeks for negotiation

Red flags to watch for

You're interviewing them too. The questions you ask matter as much as the answers you give.