AI/LLM Fundamentals for Engineers: Transformers, Fine-tuning vs RAG vs LoRA, and Agent Protocols
The fundamentals underneath every LLM system design: tokens and BPE (and why Hindi or Tamil text costs more), embeddings and cosine similarity, the Transformer and self-attention worked through with real numbers, decoder-only generation and sampling, why long context is expensive, training vs fine-tuning vs inference, prompt engineering vs RAG vs fine-tuning, full fine-tuning vs LoRA vs QLoRA with parameter and memory arithmetic, evaluation basics, MCP vs A2A, and when an agent framework beats a plain loop.
About 140 minutes · 18 exam questions · 12 interview questions · 2 code labs
What you will learn
- Tokens and tokenization: BPE, why token counts drive cost and latency, and the Indian-language tax
- Embeddings and vector similarity: cosine, dimensions, and what "semantic" really means
- The Transformer in engineering terms: from token ids to next-token probabilities
- Self-attention step by step: Q, K, V and a worked 3-token example with real numbers
- Decoder-only generation and sampling: temperature, top-p, the context window and why attention cost grows with length
- Training vs fine-tuning vs inference: how a model comes to exist and which part your team actually touches
- Prompt engineering vs RAG vs fine-tuning: a decision table for knowledge, behaviour, freshness and cost
- Full fine-tuning vs LoRA vs QLoRA: low-rank adapters, parameter counts and GPU memory
- Evaluation basics: hallucination, groundedness, correctness, and why vibes are not a release gate
- Agent protocols: MCP for tools and context, A2A for agent-to-agent delegation, and how they compose
- Agent types and frameworks: LangGraph, CrewAI, n8n-style workflows, and when a plain loop is better
- LLM glossary for system-design interviews: the 30 terms you should define in one breath
Previous: Key Technology Deep Dives: DynamoDB, Temporal & ZooKeeper