the gallery
One method at two scales. A deep case study — an agentic MLOps platform, built and measured end to end — and small free-standing prototypes, each reverse-engineering one company’s publicly-stated data-systems bottleneck and solving a slice of it, honestly measured. Same unit contract, same evidence tiers, so the reader sees one discipline applied at two densities, not a pile of demos.
coverage · the frontier is the point
A new unit must claim an open class or go materially deeper in a filled one.
Physical Intelligence · data core · class A
Catches the corruptions that leave a table structurally valid so a schema contract misses them. 6/6 vs 1/6 on a real robot dataset, zero false positives, 1.8 ms.
Read the unitfeature stores · data core · class B
Catches training data built from future feature values — the temporal leak a schema check passes and offline accuracy rewards. 0.999 leaked vs 0.771 correct; only the guardrail sees it.
Read the unitprompt caching · throughput · class C
Catches a cache that silently never hits — functional tests pass, every request pays full cost. 95% hit-ratio working vs 0% (20× redundant compute), invisible to a correctness check.
Read the unitcolumnar formats · storage · class E
A two-column query returns the right rows whether it reads 2 columns or all 12 — so correctness passes a full scan reading 44.7× the bytes. The guardrail counts bytes, not rows.
Read the unitELT · schema-drift · class G
Catches drift that keeps the column present and non-null — a silent unit change reads revenue as $157k vs a true $299k, zero errors, zero nulls. Schema/null checks pass it; the contract doesn’t.
Read the unitthe deep case study · ten units
One platform proving one claim: the classic MLOps loop still holds when the thing served is an LLM agent, once you add a token/cost axis and a reasoning-trace artifact. Explore it as a 14-stop data lifecycle.
Open the atlas