The workshop · Apps, sites, and workflows

The long tail, shipped.

Every company carries a list of things that should exist and never get built. This is the bench where they get built.

The bench

Brand sites

A wedding atelier's complete brand site, extracted from a pitch deck, live in a day, no framework.

one day to live

Internal libraries

An image database with upload, tagging, and search, engineered inside a strict platform function cap.

browser side resizing

Private AI

A local model lab proving sensitive data can stay on your own hardware: 4 billion parameters on a 6GB laptop GPU, 93 percent on a 14 trap suite.

zero cloud calls

Eval harnesses

Benchmarks that decide which model earns which job before anything ships: 22 scenarios, blind judging, more than ten models tested.

nobody grades their own homework

Inside the build

lab · local model synthetic data
» load model · 4B quantized ··· 3.9GB VRAM
» run trap suite · 14 scenarios ··· done
» hard traps ··· 14/14
» overall ··· 93%
» cloud calls made ··· 0
bench · model olympics synthetic data
representative run
Model A parse 100 · judgment 4.8 champion
Model B parse 98 · judgment 4.1
Model C parse 91 · judgment 3.2

blind judge from a second vendor

site · brand atelier synthetic data

ATELIER

weddings, curated

story · enquire

pitch deck to live site · one day

The method

Same discipline at every size: benchmarks before trust, guardrails in code, receipts on every claim, view source welcome.

Have something on the list?