LLM-Native Recommendation Systems
Understand how LLMs are replacing traditional recommendation ML infrastructure — from feature stores to prompt design, two-phase training to prefill-only serving — and what that means for building modern personalization systems
3 sections · 8 lessons
Course outline
From Feature Engineering to Context Engineering
- How traditional recommendation systems use feature engineering
- What context engineering means in LLM-native systems
- Designing compact, high-information prompts
Training LLM-Native Rankers
- Two-phase training: foundation adaptation vs. task fine-tuning
- Balancing ranking, language modeling, and alignment objectives
- Data efficiency advantages of foundation models
Serving LLMs at Recommendation Scale
- Prefill-only inference and why it matters for cost
- Infrastructure shift: feature stores to LLM serving stacks