What it is
A 10-week roadmap for learning production LLM serving by building and benchmarking one OpenAI-compatible service end to end. Guides engineers through shipping one reproducible LLM inference service, then instrumenting, load testing, optimizing, routing, and publishing benchmarks over 50 short sessions.
Key results
- 50 sessions over 10 weeks; 30 minutes each, 25 hours total
- Ends with a load harness that drives 1000+ concurrent requests
- Benchmarks FP16, FP8, INT4, speculative decoding, and KV eviction
Shows that inference-serving skills are learned faster by iterating on one production-like service than by running disconnected experiments.