Not before Stage 8
Defer- LangChain
- agent frameworks
- vector databases
- fine-tuning LLMs
- cloud certifications
- Kubernetes
- distributed training
- CUDA
- RLHF
Available from Stage 8. Scheduled week does not affect this boundary.
Scope control · Active Stage 0
What not to learn yet
Protect depth by deferring tools until the curriculum or a real project requires them.
Not before Stage 9
Defer- custom transformer implementation obsession
- diffusion models
- GNNs
- reinforcement learning depth
- computer vision specialization
Available from Stage 9. Scheduled week does not affect this boundary.
Not before Stage 11
Defer- multi-agent architectures
- five vector databases
- agent memory frameworks
- autonomous browser agents
- "prompt engineering certifications"
Available from Stage 11. Scheduled week does not affect this boundary.
Not before a project forces it
Defer- Kubernetes internals
- Spark
- Kafka
- Airflow
- dbt
- feature stores
- distributed GPU training
- custom CUDA kernels
- Terraform depth
- multiple clouds
Only when a project forces it.
Framework rule
Use the standard library / low-level concept first when educationally useful.
- learn SQL before ORM
- learn HTTP before framework magic
- learn sklearn pipelines before AutoML
- learn PyTorch training loop before high-level trainer
- learn retrieval metrics before RAG framework
- learn tool state machine before agent framework