Skip to content
KETHRA

How it works

The fitting, the eval, the handover.

How a private model goes from your data and your standard to a deployed, owned endpoint. Three steps. No ML team required.

01Step 01

Bring your data and your standard

Hand us examples of the work done well, the documents that hold your knowledge and a written note of the standard you hold. No pipeline to build, no labels to prepare.

  • A corpus of good work: documents, conversations or completions that show the standard you hold.
  • Material that carries your knowledge: internal guides, past decisions, reference texts your people actually rely on.
  • A written statement of your standard: the voice, the refusals, the things your model must always do and must never do.

You do not need to build a data pipeline, prepare labels or stand up a training environment. That work belongs to us.

kethra cli
kethra fit --data ./corpus --standard ./voice.md
02Step 02

We tailor and evaluate a private model

Kethra fine-tunes a private model on your material, aligns it to your standard and scores every version against a held-out eval built from your own bar. Nothing ships until it clears.

  • Fine-tuning on a strong open-weight base chosen for your task, trained entirely on your material.
  • Alignment to your written standard: tone, refusals and house rules are baked into the model, not added as a system prompt patch.
  • Held-out eval scores every version on voice, accuracy and refusal adherence, measured against your own examples.
  • Distillation to smaller, faster weights where it suits your latency budget or cost target.

Nothing ships until it clears the bar. You see the delta over the base model before a single request is served.

kethra cli
eval pass 47/50 · voice +31 · refusals held
03Step 03

You deploy it, and you own it

One command puts the model behind a private endpoint. The weights are yours, exportable, on our infrastructure or inside your own. Refinement continues as your standard moves.

  • One command deploys the model to a private, OpenAI-compatible endpoint. Your engineers swap a base URL and a model name.
  • The weights are exportable in a standard format, hosted on our infrastructure or entirely inside your own perimeter.
  • Versioning and rollback are built in: a new version never silently replaces a known-good one.
  • Refinement runs on a regular cycle as your standard moves, so the fit stays true rather than drifting out of shape.

Your data trained your model. The model belongs to you. Both facts are stated in the contract and do not change.

kethra cli
kethra deploy --private → wss://api.basisresearch.tech/v1
Ownership

What you keep

01

The tailored weights are your property, in writing.

02

Export them at any time in a standard format.

03

Host on our infrastructure or entirely inside yours.

04

Your data trains your model only, never the shelf.

A tailored model is not a subscription to someone else's shelf. The weights are yours to export, to host, to keep behind your own walls, and to carry with you if you ever leave. Kethra does the tailoring. The garment is yours.

Questions

Asked often

A model made to measure.

Start free and tailor a sample in a minute, or book a fitting and we will build you a private model of your own.

Start freeBook a fitting