I build AI systems that hold up.

A modern filesystem for removable flash, built by cutting every feature a general-purpose filesystem carries that an SD card never needed — then verifying what remained until it could not reasonably be doubted.

• Five invariants govern every write: race-to-idle, fill-or-kill, ordered commit, CRC32C only, no data journaling
• A human architect set the constraints; AI implemented narrowly against them
• Seven independent AI models audited the finished code adversarially
• 310 crash-injection trials and 17/17 xfstests MUST_PASS — reproducible and unredacted

Design decisions were forced through named, demanding users — a wildlife photographer shooting 7,500 frames a day, a videographer recording 8K sustained, and a deliberately hostile kernel reviewer who accepted no claim without a log.

Traditional Bible translation is a multi-year, sequential process gated by scarce expert time. GOI Bible treats it as software: translate the smallest verifiable unit, check it independently, route human judgment only where the checks raise a flag, and make the whole run repeatable.

• AI drafts one verse at a time from tagged Hebrew and Greek source text
• Independent checks test nouns, clauses, negations, and numbers
• Flags route scarce human review to exactly the places that need it
• Versioned and auditable — a correction re-runs the pipeline instead of restarting the project

Published in English, Simplified Chinese, and Traditional Chinese, with a full offline SQLite distribution and a global verse ordinal for programmatic use.

Pareto says 20% of inputs create 80% of results. 64-4 takes it one level deeper: roughly 20% of the important 20% produces about 64% of the total result. The goal is never to do more — it is to find the few decisions, constraints, advantages, and points of friction that disproportionately control the outcome.

CoreCut4 is the consulting arm: the same four lenses applied to real AI, operations, and systems problems, turned into workflows, tools, and decisions a leadership team can act on.

A working answer to the question every regulated industry asks about AI: how do you use a language model against a system of record without ever letting it touch the record?

• Human narrative is semantically compiled into a strict, versioned ingest contract
• AI never writes SQL and holds no mutation authority — it proposes, it does not commit
• Human approval gates every database write; approved events append immutably
• Replay reconstructs truth from the event log; every view is a projection, never authoritative
• Ambiguity and conservation failures surface before approval, not after

Fifteen explicit contracts, an ontology registry, and an audit stream — built as a deliberately small prototype in PHP and SQLite to prove the governance model rather than the tech stack.