Notes from one level down

Data and AI engineer. I write about the machinery under the tools, how to build AI systems that hold up, and sometimes about making music with a model in the room. Slow reads, on purpose.

systems

Your Spark job is a mailroom The shuffle explained with clerks, slips, and postage. Once you see the mailroom, stage boundaries, skew, and half of your tuning options become obvious. published 2026-08-05 · updated 2026-08-08 · 9 min · systems The disk has two speeds Sequential and random are not two workloads. They are two different machines hiding in one box, and your access pattern picks which one you bought. published 2026-08-05 · updated 2026-08-08 · 6 min · systems Exactly once is a habit, not a feature Checkpoints, idempotent writes, and why the honest phrase is effectively once. The discipline generalizes from streams to agents without changing shape. published 2026-08-05 · updated 2026-08-08 · 6 min · systems Put like near like Z-ordering, vector indexes, and cache lines are one instinct wearing three costumes: arrange data so that things wanted together sit together. published 2026-08-05 · updated 2026-08-08 · 5 min · systems The most expensive line is the one you never wrote No line of your code says shuffle, or garbage collect, or page fault. A field guide to the invisible costs that own your latency budget. published 2026-08-05 · updated 2026-08-08 · 6 min · systems

applied ai

Prompts are contracts Treat prompts like schemas: versioned, tested, with named failure modes. The single mindset shift that separates demos from systems. published 2026-08-05 · updated 2026-08-08 · 6 min · applied ai The retrieval ladder Dense, hybrid, reranked, filtered. Four rungs, each one patching a named blindness of the rung below, and a yardstick you build before you climb. published 2026-08-05 · updated 2026-08-15 · 6 min · applied ai An agent is a consumer of its own event log Checkpoint, resume, replay. The reliability tricks agents need were invented by stream processors years ago, and they transfer cleanly. published 2026-08-05 · updated 2026-08-08 · 6 min · applied ai Evals or it didn't happen Golden sets, judges with named biases, and gates that fail builds. How claims about AI behavior become measurements instead of vibes. published 2026-08-05 · updated 2026-08-08 · 6 min · applied ai The KV cache is the new working set Why serving LLMs is a memory real estate problem, what batching actually buys, and how paging ideas from the 1960s came back to run 2026. published 2026-08-05 · updated 2026-08-08 · 7 min · applied ai

the practice

Learning one level down A method for depth: pick the layer below your job, descend nightly in small artifacts, and let the diagrams do the remembering. published 2026-08-05 · updated 2026-08-08 · 6 min · practice How to ride a paid cohort Courses give you sessions. A sidecar of prep, labs, and teach back notes is what turns sessions into capability. Template included. published 2026-08-05 · updated 2026-08-08 · 5 min · practice Interview stories are systems stories How to narrate technical work honestly: mechanisms first, numbers you can defend, and the difference between naming an ocean and having been under it. published 2026-08-05 · updated 2026-08-08 · 6 min · practice

studio notes

Producing with a model in the roomreading copy I am building the synth I make music with, codename syn1, and the model has three jobs in that room: bench partner, menu widener, borrowed ears. Where the line runs, from the build log. published 2026-08-05 · updated 2026-08-08 · 7 min · studio What mixing taught me about evalsreading copy Reference tracks are golden sets and tired ears are a biased judge. Then it runs the other way: my synth carries a three million assertion regression suite, bench card included. published 2026-08-05 · updated 2026-08-08 · 8 min · studio

fifteen essays, all published here, open to anyone, no login. four of them name a deeper private piece that is not part of this site. the two studio notes are reading copies: their published home is the lab's research shelf. new pieces land here as they are written.