What shipping five small AI repos taught me

I decided the fastest way to show I can do AI engineering was to build the pieces, not describe them. Five repos, each small, each runnable, each with tests and CI. No fabricated history, no empty commits.

• rag-pipeline: chunk, embed, FAISS, retrieve, grounded answers with citations
• llm-finetune-lab: LoRA and PEFT fine-tuning with prompt masking
• dynarq-agent: a tool and skill interface with a tool-calling loop
• llm-eval-harness: deterministic metrics plus an optional LLM judge
• dynarq-shield: a prompt-injection and data-leak firewall

The lesson that stuck: quantity of repos is not the lever. Depth and real-world signal are. One repo that runs end to end says more than ten that only have a README.