New Odyssey

Outcomes · For engineering teams

Lower model cost without losing accuracy

A smaller open model is fine-tuned on your own examples and measured against your current model, so you switch only if it holds up.

Who it is for

Who this is for

AI and ML engineering teams running a repeated task in production, such as ticket classification or structured extraction, on a large general-purpose model.

What changes

Before and after

Before

  • A narrow, repeated task runs on a large general-purpose model.
  • Whether a smaller model would do the job is a guess.
  • Any saving is a projection nobody has measured.

After

  • Your current model is measured on held-out examples against acceptance criteria you set.
  • A candidate open-weight model is compared with that baseline on task accuracy, structured-output validity, latency and estimated cost.
  • You deploy the candidate, export the supported model artefacts, or reject the experiment.

How it is delivered

Delivered by a product, not a project

Delivered by Daidon. It is in development and not available to run today.

  • Prepare: upload CSV or JSONL examples; Daidon checks quality, duplicates and sensitive data, and splits training, validation and held-out test sets.
  • Train: a managed supervised fine-tune (LoRA where supported) on an open-weight model, starting with selected Qwen models.
  • Every training record keeps its provenance.

What you provide

What we need from you

  • Examples of the task in CSV or JSONL, from your own production use.
  • The acceptance criteria the new model has to meet.
  • Your current model, as the baseline to beat.

How the result is measured

How you will know it worked

  • On a held-out test set: task accuracy, structured-output validity, latency and estimated cost, candidate against baseline.
  • Projected savings are estimates until a reproducible experiment measures them.

How to start

How to start

Request early access

Daidon is in development. Requesting early access puts you on the list; it does not start a trial.

Good to know

Limits, stated plainly

  • "Without losing accuracy" is the test, not a promise. If the candidate does not clear your bar, the answer is to reject it.
  • Fine-tuning is not always the answer. Knowledge that changes often is usually better served by retrieval, better prompts or ordinary code, and Daidon will say when training is not worth it.
  • Unverified model outputs never become trusted training labels automatically.
  • There is no SDK, install command or hosted service yet.

For engineering teams

Other outcomes