Step 01
Step 1
Open /models; filter by stage, modality, or audience.
ComplianceRetail + B2B· marketing-claim · models-in-catalog
Public model cards.
Every model Mukodi runs has a public card — stage, license, training data, limitations.
The problem
Why this exists.
Most AI companies treat their model inventory as a trade secret. Regulators, journalists, and curious users get a paragraph in a marketing page.
We want every model in the catalog to have a public card with the same depth on day one — stage, license, foundation, training-data summary, drift status, version history.
How it works · 3 steps
From input to outcome.
- Step 02
Step 2
Tap any card; see the full model spec including limitations and version history.
- Step 03
Step 3
Tainted or replacement-in-progress models carry an explicit caveat — we do not hide them.
Under the hood
The models and routes behind it.
Models called
No models cite this feature directly — backing logic is deterministic Python.
API routes
- GET /api/v1/model-cards
Compliance posture
- Model cards are generated from the same catalog YAMLs the runtime reads — no marketing-side rewriting.
- Apache 2.0 / MIT license filter ships at the top of the page; tainted is marked vermilion.
- Boot-time drift guards lock the catalog so cards cannot silently drift from runtime.
What it's not
Model cards summarise; they do not replace evaluation harnesses or DPIAs. The full risk record lives in our internal compliance documents.
Related · features & stories