#costengineering

Two runs died on rate limits in the same week. Not because the work was too big. Because the most...

A young AI outpainting tool that turns 9:16 video into 16:9 starts handling real paid work: vendor retries instead of abandoned renders, 720p quietly billed as 1080p, and a two-tier fallback chain across models that disagree with each other's documentation.

Collapsing four separate ad-creation entry points into one brief-driven flow, building a VSL Generator and an AI Presenter mode on top of it, and the sample-rendering and cost-quoting accuracy work needed to make the pre-review sample trustworthy before a paid run.

Replacing a Drive-folder-scan run list with a fail-open Supabase index, building a durable per-charge cost ledger after finding the delta was computed and thrown away, and the 1,897-binding lint sweep that found four real bugs hiding in dead code.

Auditing a video generation voice pipeline: billing what vendors actually charge, a budget/quality SKU ladder that opened backwards, content-keyed voiceover caching, and a pronunciation dictionary nobody could hear.

Eight PRs closing out a QA pass on an AI presenter tool: a flaky call that silently disarmed a spend gate, an aspect-ratio bug that reproduced on every run, paying to render pixels and then binning them, and diagnostic logs that a routine redeploy erased.

Real money billed at $0, a resume path that silently re-decided a paid render's settings, a $0.82 measurement of the wrong build, and a reconciliation script that had to recount artifacts because nobody had vendor dashboard access.

Packaging a render-six-references-and-measure workflow into a reusable skill for an AI ad generator, and the audit backlog it surfaced: dead flags, a stop-ship brand gate, a misspelled brand name on a finished master, and a documentation culture that corrects its own wrong claims

I watched my email extraction pipeline burn money re-deriving the same fields over and over—then I realized the bug wasn’t “missing cache,” it was ...

Uniform frame sampling made my screen-workflow analyzer both expensive and blind to short UI bursts. I replaced it with an adaptive sampler that sc...