AI
Technical notes from the AI domain: from machine learning fundamentals and deep learning internals to LLM application development (Prompt / RAG / Agent / Fine-tuning), continuously updated along with learning.
Collected
- Machine Learning Fundamentals — Math foundations and classic algorithms
- Deep Learning — Neural networks, CNN / RNN / Transformer
- Prompt Engineering — Prompt design and evaluation
- RAG and Agent — Retrieval-augmented generation and agent applications
- Model Fine-tuning — LoRA / QLoRA and deployment
Planning Checklist
| Topic | Description | Status |
|---|---|---|
| Machine learning fundamentals | Linear algebra/probability & statistics crash course, regression/classification/clustering, evaluation metrics | Planned |
| Deep learning | Backpropagation, optimizers, CNN/RNN/Transformer architectures | Planned |
| PyTorch | Tensors, automatic differentiation, model training pipelines, distributed training | Planned |
| LLM internals | Attention mechanism, pre-training and alignment (RLHF/DPO) | Planned |
| Prompt engineering | Prompt patterns, structured output, evaluation methods | Planned |
| RAG and Agent | Vector retrieval, tool calling, multi-agent frameworks (LangChain/LlamaIndex) | Planned |
| Model fine-tuning | LoRA/QLoRA, data construction, evaluation and deployment | Planned |
| AI engineering | Inference optimization, quantization, serving (vLLM/Ollama) | Planned |
| Advanced RAG | Index structures, advanced retrieval, reranking, GraphRAG/Agentic RAG, evaluation | Planned |
| LangGraph orchestration | Graph state machine, checkpointers and stores, interrupt-based HITL, production practices | Planned |
| GUI agent (computer use) | Perception, action space, verification loop, sandboxing, benchmarks and limits | Planned |
| Platform & framework selection | Stack layers, route comparison, decision framework, TCO, avoiding lock-in | Planned |
Writing Guidelines
- Theory notes should include formulas along with intuitive explanations, avoiding formulas without examples;
- Practice notes should provide runnable Notebook or command-line examples, stating dependency versions;
- Concepts related to papers should link to the paper notes in Literature Collection;
- File names use kebab-case, and new files should be registered in the
config.mjssidebar.