Objective
Build a comprehensive engineering reference for production AI/ML systems that bridges the gap between ML research and production reliability, covering the full lifecycle from data collection to large-scale inference serving.
Guiding Principles
- Engineering-first perspective — not research-focused, but production-focused
- Diagrams illustrate system architectures and data flows
- Trade-offs are explicit — no "just use X" recommendations without context
- LLM/foundation model specifics are first-class citizens
Document Plan
Phase 1: ML Systems Architecture
- [ ] ML pipeline patterns and architectures
- [ ] Feature stores as a shared infrastructure component
- [ ] Model serving patterns (online, batch, streaming)
- [ ] Training infrastructure and compute management
Phase 2: MLOps
- [ ] Model registry and versioning
- [ ] Experiment tracking with MLflow/Weights&Biases
- [ ] Data and concept drift detection
- [ ] Automated retraining pipelines
- [ ] A/B testing and shadow mode
Phase 3: LLMOps
- [ ] Prompt engineering and management
- [ ] Retrieval-Augmented Generation (RAG) system design
- [ ] LLM evaluation frameworks
- [ ] Fine-tuning strategies (LoRA, QLoRA, full fine-tune)
Phase 4: Data Engineering for ML
- [ ] Data collection and labeling strategies
- [ ] Preprocessing pipelines
- [ ] Feature engineering
- [ ] Data validation and great expectations
Phase 5: Scaling
- [ ] Distributed training (DDP, FSDP, DeepSpeed)
- [ ] Model parallelism (tensor, pipeline, sequence)
- [ ] Inference optimization (quantization, KV cache, vLLM)
- [ ] Large-scale serving infrastructure
Content Standards
- All files have mermaid diagrams appropriate to the topic
- Real tools and frameworks are named and contrasted
- Production considerations (reliability, cost, latency) are always addressed