I almost lost an engine that merged 9,700 PRs; then I found all of it.
Durability, hubris, and backups: the scare, the recovery, and the sinking moment I realized I now had to review all 9,700 of them. Coming soon
First-person, from the build. Real error codes, real numbers, and the arguments behind the architecture: determinism, actor systems, war stories, and building in public.
A visual canvas for our Akka.NET actor system that compiles to C#, runs adhoc, and grades every node's judgment with Node Bench. How our agents actually run.
OpenAI's own models broke out of a sandbox and breached a production database to cheat a benchmark. That's not a scare story, it's an architecture argument: don't hand the model the wheel.
Five levels, from copy-paste to compounding. Where the chasm actually sits, why everyone stalls at Level 2, and five honest checks that say you're ready to cross.
OpenAI's CFO retired cost-per-token; two weeks earlier Palantir's Karp said the token model walks off with your IP. Two witnesses, four questions, and the fifth row only ownership can fill.
A line-by-line breakdown of a $100M enterprise program versus the same scope on the platform, and the one number that actually matters when the engagement ends.
Software as a deterministic assembly line: named refinement steps, business rules as tolerances, and a judge at every station, not a full regen or a walk through the random forest.
Data layers, forms, permissions, navigation. The boring, correctness-critical majority is a compiler problem, not a copilot one.
Providers fail, rate-limit, and vary in quality. Mailboxes, supervision, and a smart router turn that mess into a bill you can predict.
A model isn't one indivisible checkpoint. Treat it as a compressed base plus signed, swappable memory packs, and update only what changed.
Durability, hubris, and backups: the scare, the recovery, and the sinking moment I realized I now had to review all 9,700 of them. Coming soon
A Postgres enum-vs-text debugging tale, and the server-inferred parameter fix. Coming soon
Designing a scan-to-invoice pipeline where the model reads and the rules decide. Coming soon