Skip to content
leeveel

Writing · since 2024

Notes from the studio.

Things we've learned. Things we've changed our minds about. Drafts in public.

Filter
★ Featured · Nº 013 · Jun 305 min read

The QA list nobody wants to read

91% of the bugs that reach a real user are off the happy path everyone tested. So before every ship we run the QA list nobody wants to read — empty states, double-submits, offline mid-action, names with an apostrophe. Here it is, in order.

by Ria P.#process
Read on →

Archive

12 posts
  1. 012

    We trained the model in-house

    A recommender doesn't need a foundation model. We shipped one for a hospitality client — a few thousand parameters, retrained every Monday on a single CPU. Most 'AI feature' briefs we see are a ranking problem in a costume.

    Thanos K.6 min#ai
    Jun 24
  2. 011

    The spreadsheet is the spec

    Before we replace a client's spreadsheet with a real internal tool, we read every column, every macro, and every dead tab. The spreadsheet already is the system — the rebuild is a transcription with permissions and a query layer.

    John P.6 min#custom-systems
    Jun 23
  3. 010

    Most of your scan findings are noise

    A passing scan and a defensible posture are different things. Here's how we tell the noise from the six findings that matter.

    Thanos K.6 min#security
    May 26
  4. 009

    Half the time, you don't need a mobile app

    About half the briefs we see don't need an app. Here's the question we ask first.

    John P.5 min#mobile
    May 10
  5. 008

    Why we ship on Fridays

    The week's worst day to ship turns out to be the most honest one. We argue why, then show our deploy log.

    Nephele K.4 min#process
    Apr 18
Showing 5 of 12