Labs & capstones

Curriculum labs, interactive tools, and capstone projects for inference engineering.

Observable mini inference engine: Build a small serving engine with a paged KV allocator, continuous batching, chunked prefill, preemption, and prefix sharing, then make every scheduler decision replayable from a trace. It proves you understand the state machine that vLLM and SGLang are built on, not just their flags.

Reproducible benchmark laboratory: Build an endpoint qualification harness that validates output semantics, drives open-loop and closed-loop load, records every request and error, and captures a run manifest. It proves you can produce a latency number another engineer can reproduce and trust.

Hardware and model fit planner: Build a decision artifact that takes exact GPU SKUs, weight formats, runtime overhead, KV capacity, and a target SLO and returns a GPU count, a cost per million tokens, and an uncertainty band. It proves you can turn arithmetic into a purchase recommendation and then check it against a measurement.