Go from using LLMs to understanding and shaping them: see the forward pass and the KV cache, fine-tune with LoRA/QLoRA on a free GPU, quantize and serve efficiently, and evaluate and secure what you ship. This is the line between "I call an API" and "I build LLMs." Builds on your BYO-7 tokenizer and BYO-8 mini-GPT.
One end-to-end LLM-engineer portfolio piece: QLoRA fine-tune an open model on a domain dataset (security Q&A or finance) → quantize it to GGUF → serve it (Ollama/vLLM) → evaluate before/after with lm-eval-harness + an LLM-as-judge rubric → red-team it against the OWASP LLM Top-10 (BYO-16). One write-up ties it together — the standout "LLM engineer" resume item most candidates lack. Read a key paper (LoRA or DPO) with your 3-pass method. Ready for Phase 7 when you can explain a transformer end-to-end, and fine-tune + quantize + serve + evaluate + secure an open model.
Full guide: phase-6b-llm-indepth.md ↗