career

kicking off theplatform - a new system, a small language, and three ideas to fit into one runtime.

three months in. why arity-classed vector evaluation still gives modern systems a useful design challenge.

the language has a name. settling the dispatch rule before the syntax.

a layout note I keep having to reread. why columnar wins by 10x once two passes are involved.

the only discipline that has stayed put after a year of writing the o kernel.

a notebook entry from work i cannot share publicly. why grammars-as-values landed for me on this protocol first.

when grammars become first-class, the language stops being two languages.

channel selection works until a transaction needs values from two channels. the reaction has to be the unit.

the first reactor is in. join patterns provide atomic matching that actor mailboxes do not.

the bet i made in 2016 that i am, finally, fairly sure was right.

composing a socket, a parser, and a reaction across subsystem lines, with no boundary to pay for.

three laziness modes for join reactions. the runtime knows them. the surface still does not expose them.

tachyon's workspace is one crate per arity class. the cleanest split i have tried.

an audit of every peg grammar shipped under std/. the small line count is evidence for grammars-as-values.

ten years of theplatform. what i think now that i did not think then.

i needed an mpmc channel with async support. hyperbridge separated that concurrency problem from theplatform.

two days trying to wake n receivers in order. one commit message that is more honest than i would normally write.

two months of using hyperbridge in theplatform. one bug class i had been ignoring is now closed.

the head-tail rewrite. one flag bit makes the difference between "spin once and recover" and "lock to be safe". 0.2.5 ships tonight.

four years in production. an embeddable, smaller, c-shaped sibling has been quietly forming.

a year and a half of thinking. the database i want to ship lives in someone else's binary, so it has to be in c.

the rayforce repo went on disk today. seven files, a hundred and three lines, no allocator, no test.

the kernel signature on day one is the kernel signature today, even though the discipline around it changed.

a year in. writing down which of k's choices made it into rayfall and which i deliberately left behind.

a week reading kparc's annotated minimal k. two ideas i stole and one i did not.

a u64 bitmap, a mask, a count-trailing-zeroes. the smallest fitting order in two instructions.

every task carries its own allocator. cross-thread frees never touch a lock.

the repo is still private. i am writing the python orm anyway. it is the cheapest way to find out which parts of the c api are wrong.

a single c file, a makefile, and a csv. the rule that emerged after four months of measuring.

five years of thinking, two and a half years of private commits. today rayforce is open source.

seven days, four desktop ui frameworks, and the decision to keep the database UI close to the engine.

axldb is the rayforce kernel with the database parts removed to test where the engine boundary actually sits.

four months, one hundred and seventy-nine commits, then closed. what the experiment gave back and what it did not.

the parallel research project and the production system are now one repo. keeping the language. swapping the engine.

rayforce2 with the teide pipeline underneath rayfall. morsels make cancellation, fusion, and profiling structurally simpler.

the experiment teidedb is built on. eighteen years of intuition into a spec dense enough that something else can deliver the implementation.

a new query surface landed this week. eav triples, semi-naive evaluation, and a fixpoint that fits the morsel pipeline naturally.