Lilly Tech Systems
NLP Interview Landscape
The NLP hiring landscape has shifted dramatically since 2022. Classical NLP skills (regex, rule-based systems, feature engineering) still matter, but interviews now heavily emphasize transformers, LLMs, and production deployment. This lesson maps the terrain so you know exactly what to prepare for.
How NLP Interviews Have Changed
Before 2020, NLP interviews focused on feature engineering, statistical methods, and classical ML pipelines. Today, the expectation is fundamentally different.
| Aspect | Classical NLP (Pre-2020) | Modern NLP (2022–2026) |
|---|---|---|
| Core Knowledge | TF-IDF, n-grams, POS tagging, dependency parsing, CRFs | Transformers, attention, BERT/GPT, prompt engineering, RAG |
| Model Training | Train from scratch on task-specific data | Fine-tune pretrained models, few-shot/zero-shot, LoRA/QLoRA |
| Coding Questions | Implement tokenizer, build Naive Bayes classifier | Write prompt templates, implement RAG pipeline, use HuggingFace |
| System Design | Build text classification pipeline | Design LLM serving infrastructure, RAG architecture, multi-model routing |
| Evaluation | Accuracy, F1, precision, recall | BLEU, ROUGE, BERTScore, human evaluation, LLM-as-judge |
| Production Skills | Feature stores, batch inference | Token cost optimization, context window management, guardrails, latency budgets |
NLP Role Types and What They Test
Different roles emphasize different skills. Understanding which role you are interviewing for lets you focus your preparation.
NLP Research Scientist
Focus: Novel architectures, training methodology, scaling laws, paper reproduction. Expect deep questions on attention mechanisms, positional encoding, training stability, and ablation studies.
Companies: Google DeepMind, Meta FAIR, OpenAI, Anthropic, Microsoft Research
NLP/ML Engineer
Focus: Building production NLP systems. Fine-tuning models, data pipelines, evaluation, deployment, monitoring. Expect system design and coding rounds alongside ML theory.
Companies: Google, Amazon, Meta, Apple, Netflix, Uber, Airbnb
LLM/GenAI Engineer
Focus: Building applications on top of LLMs. RAG, prompt engineering, agent frameworks, guardrails, cost optimization. The newest role category with the fastest-growing demand.
Companies: Startups, enterprise AI teams, consulting firms, all Big Tech
Applied Scientist
Focus: Bridging research and production. Design experiments, run A/B tests, improve models iteratively. Expect both paper discussions and production-oriented system design.
Companies: Amazon, LinkedIn, Spotify, Pinterest, Salesforce
Typical Interview Format
Most NLP interviews at top companies follow this structure across 4–6 rounds:
| Round | Duration | What They Test | How to Prepare |
|---|---|---|---|
| Phone Screen | 45–60 min | NLP fundamentals, basic coding, motivation | Review Lessons 2–4 of this course. Practice explaining concepts clearly in 2–3 minutes. |
| Coding Round | 45–60 min | Implement NLP algorithms, use PyTorch/HuggingFace, data processing | Practice tokenizer implementation, text preprocessing, simple model training loops. |
| ML/NLP Deep Dive | 45–60 min | Transformer internals, training strategies, evaluation, recent advances | Review Lessons 3–5. Be ready to whiteboard attention computation and loss functions. |
| System Design | 45–60 min | Design NLP systems at scale: search, chatbot, content moderation | Practice end-to-end design: data pipeline, model serving, monitoring, cost analysis. |
| Behavioral | 30–45 min | Past projects, conflict resolution, leadership, ambiguity handling | Prepare 5–6 STAR stories from NLP projects. Quantify impact (latency reduced 40%, accuracy +5%). |
What Companies Actually Want
Based on hundreds of interview debriefs from FAANG and top-tier companies, here is what separates "hire" from "no hire" candidates:
- Depth on transformers: Can you explain multi-head attention from first principles? Not just "BERT uses attention" but the actual Q/K/V computation, why we scale by sqrt(d_k), and how positional encoding works.
- Production mindset: You do not just train models — you think about latency, cost, monitoring, failure modes, and data drift. You know the difference between a Jupyter notebook demo and a production system.
- Trade-off reasoning: When asked "should we fine-tune or use RAG?", you do not give one answer. You ask about data volume, latency requirements, update frequency, and cost constraints before recommending an approach.
- Current awareness: You know about recent developments: mixture of experts, long-context models, multimodal LLMs, reasoning models, and can discuss their practical implications.
- Clear communication: You can explain complex NLP concepts (attention, RLHF, BPE tokenization) to a senior engineer who is not an NLP specialist. Clarity beats jargon.
Preparation Strategy
Here is a structured 3-week plan to prepare for NLP interviews using this course:
Week 1: Foundations
Complete Lessons 1–3. Focus on tokenization, embeddings, and transformer architecture. Practice explaining each concept out loud in under 3 minutes. Write code for BPE tokenization and attention computation from scratch.
Week 2: Applications & LLMs
Complete Lessons 4–5. Study NLP tasks (NER, classification, summarization) and LLM-specific topics (RAG, RLHF, prompt engineering). Do 2 mock interviews focusing on Q&A format.
Week 3: Practice & Polish
Complete Lessons 6–7. Work through practical challenges and rapid-fire questions. Do 2 full mock interviews under time pressure. Review weak areas and refine your STAR stories.
Key Takeaways
- Modern NLP interviews focus 70% on transformers/LLMs and 30% on classical foundations
- Know which role type you are targeting — research scientist, NLP engineer, LLM engineer, or applied scientist
- Companies want depth on transformers, production mindset, trade-off reasoning, current awareness, and clear communication
- Follow the 3-week preparation plan: foundations, applications, then practice under pressure
- Practice explaining concepts out loud — reading is not enough to pass an interview