Colour tokens
Ground
ground
#F8F4EC
ground-alt
#F0EAD8
ground-deep
#E5DCC8
Ink
ink
#1C1208
ink-mid
#3D2C18
ink-muted
#7A6548
ink-ghost
#B8A88A
Amber -- the marigold flower
amber
#C07A0A
amber-mid
#A86308
amber-deep
#7A4400
amber-lt
#FDE9B8
amber-pale
#FEF5DF
Forest -- oak leaf
forest
#2D5016
forest-mid
#3D6B20
forest-lt
#C8DFB0
Wood -- walnut to oak
wood-dark
#2C1810
wood-mid
#5C3420
wood-oak
#8B5E3C
wood-lt
#C49A6C
Typography
Marigold ⚘
Model weights are the fat layer. Workflows are thin clients over a shared substrate. The eval is the task specification.
All computation is model inference.
Inference API
Typed operations over self-hosted open-weight models. One container image, one EFS weight cache, per-model isolation. The marginal cost of adding a model falls with each one added.
What runs on it
Private model hosting. Workflows. Evals.
01
Syntactic compatibility and semantic compatibility are separate concerns. The system enforces the former and is silent on the latter.
Components -- light ground
Buttons
Tags
Ornament divider
Pillar grid
01
Inference API
Async, typed, multi-modal. Text, image, audio, depth.
02
Workflow execution
Declarative YAML pipelines. Workflows are thin clients.
03
Evals
The eval is the task specification. Data plus labels plus model.
Cards
Document Q&A
Extract, answer, and embed document content in one pipeline.
Accessibility audio
Alt-text and audio descriptions in multiple languages from one image.
Components -- dark panel (wood-dark)
Tags on dark
Typography on dark
Production AI pipelines
done properly.
Model weights are the fat layer. Workflows are thin clients. All computation is model inference.
Pillar grid on dark
01
Prototype
Validate AI concepts rapidly in realistic conditions.
02
Production
Secure infrastructure and robust cloud deployment.
03
Scale
Optimise for growth and operational reliability.
Decorative patterns
Section header strip -- amber rule
Section border -- oak dash
Both patterns are CSS-only. The amber strip marks top-of-page and major section breaks. The oak dash marks sub-section boundaries where a full rule is too heavy.
Core vocabulary in use
These phrases should recur across product copy, articles, and documentation. See TERMINOLOGY.md.
Model weights are the fat layer.
Value and generality sit in the weights, not in the applications above them. The foundation model is infrastructure.
Workflows are thin clients.
An application that calls a model does not own the capability it is using. It is a declaration of intent over a shared substrate.
The eval is the task specification.
The labelled dataset is the formal statement of what "correct" means for a pipeline. Not a downstream quality gate -- the primary artefact of any engagement.
All computation is model inference.
No regex branches, no hard-coded rule sets, no custom classifiers. Every decision in a pipeline is the output of a model that can be inspected, replaced, or retrained.
Data plus labels plus model equals evals.
The compact form. Useful as a pull-quote or a sentence that stands alone. Avoid writing it as a formal equation outside technical contexts.