Fresher / New Grad

Start with the basics: what are transformers, attention, and embeddings? Build conceptual intuition.

• Start hereExplain scaled dot product attention.
• Explain why transformers replaced RNNs for language modeling.
• Which components make up a standard transformer block?
• Why is cosine similarity preferred over Euclidean distance for text embeddings?
• What is the primary purpose of embeddings in modern NLP?
• Fill in the scaled dot product attention formula.

Learn about RAG, vector databases, and tokenization, the building blocks of modern AI apps.

• Start hereWhich metric best measures whether a RAG answer is grounded in the retrieved context?
• When should you fine-tune instead of using RAG?
• What does RAG primarily help with in LLM based applications?
• Why combine BM25 with dense embeddings for retrieval?
• HNSW vs IVF, when…
• Match each vector index type to its primary characteristic.

Survey the landscape: agents, fine-tuning, evaluation. Start drilling questions across all formats.

• Start hereWhat is RLHF, and why is it used after pretraining?
• What does LoRA do, and why is it popular for fine-tuning?
• Which metric best measures whether a RAG answer is grounded in the retrieved context?
• Design a RAG evaluation harness, what do you measure and how?
• When should you fine-tune instead of using RAG?
• PPO vs DPO, what's the practical difference?

Focus on system design basics, inference optimization, and mixed-format practice under time pressure.

• Start hereWhat is the KV cache in transformer inference?
• In LLM serving, what is the primary driver of end to end latency for a generation request?
• Design a RAG evaluation harness, what do you measure and how?
• Why combine BM25 with dense embeddings for retrieval?
• Estimate the per call cost of a typical RAG chatbot using GPT-4o-mini.
• Spot the error in this explanation of temperature.