• How LLM works

Tokenization

Context Window

Transformers

Decoding

Embedding Model

Hallucination

LLM Pipeline

This is what happens when a user sends a message and gets a response. Every concept you learned plugs in here.

• Flow Diagram
• Walking through one specific example
• How Models Are Trained
• Why training matters for FDE

Example Models (the landscape you should be able to name)

TRAINING (one-time, by Google/OpenAI/Anthropic)
                  ════════════════════════════════════════════════
                  Pretrain → SFT → RLHF/DPO → (optional fine-tune)
                  ════════════════════════════════════════════════
                                      │
                                      ▼
                              DEPLOYED MODEL
                                      │
                                      ▼
                  INFERENCE (every request, by you)
                  ═════════════════════════════════
                  Tokenize → Embed → Position → Transformer blocks
                       → Final layer → Logits → Decode → Detokenize
                  ═════════════════════════════════
                                      │
                                      ▼
                      WRAPPED IN A SYSTEM (your job)
                      ══════════════════════════════
                      RAG (retrieval) + prompting + agents + 
                      eval + observability + security + scaling
                      ══════════════════════════════

Three time horizons:

• Training: weeks-months, billions of dollars, done by labs
• Inference: milliseconds-seconds per request, you pay per call
• System: the architecture around the model, where FDE work lives

Your job as an FDE is mostly in the third layer. But you need to understand the first two to make good choices in the third.