Kaiyin Zhong — Data & AI platforms that hold their shape

Essays

llmops

Lazy by design: what MCP can learn from -h

A CLI costs zero tokens until the model asks for help, then reveals itself one -h at a time. MCP hands over every tool schema before the first question is asked. The protocol has five escape hatches (pagination, listChanged, meta-tools, deferred loading, code execution) and every one of them is a way of teaching MCP to behave like a shell.

Jul 23, 2026 · 5 min

A key wrapped in a key wrapped in a key Jul 21, 2026 You type one passphrase, but ZFS native encryption juggles three kinds of key: a wrapping key derived from what you know, a random master key it guards, and a per-dataset data key derived from that. The indirection isn't bureaucracy — it's what makes changing a passphrase instant, inheritance sane, and 'send an encrypted backup to a machine you don't trust' a one-flag operation. Here's the whole hierarchy, the GCM authentication underneath, why copy-on-write lets ZFS reuse a per-block IV safely, and an honest account of what the encryption does not hide.
Two disks down, still standing Jul 21, 2026 The obvious way to survive two dead disks — add a second parity column — fails for a reason worth understanding: a second XOR is the same equation twice. RAIDZ2's fix is to compute its second parity in a finite field, GF(2⁸), weighting each disk by a distinct power of a generator. That turns two dependent equations into an always-solvable 2×2 system. Here's the field, the syndromes, the full reconstruction, and a playground where you can take out any two disks and watch the algebra put them back.

Data, engineering, architecture. A decade of building data and ML platforms — research pipelines, production ML, and now LLM infrastructure as CDO at Swoep. I care most about the layer people skip: the foundation that decides whether everything on top of it actually works.