Tailor — Customisation and fine-tuning
A general model does not know your products, your terms or your exceptions. Onega customises AI to your work in the lightest way that passes your tests: better instructions and retrieval first, fine-tuning when the evidence says it pays, and a smaller model on your own hardware when cost, latency or data demand it.
01 — The ladder
Each step costs more to build and to maintain than the one before, so we climb only when the evaluation set shows that the previous step is not enough.
Prompts, output schemas, examples and guardrails tuned to your cases. Quick to change; no training.
Answers grounded in your documents and records through embeddings, search and reranking, with sources shown.
Open-weight models adapted on your approved examples with parameter-efficient methods, for tone, format, classification or domain language.
A smaller model trained to match a larger one on your task, so it runs on hardware you own at a cost you can plan.
02 — Evaluation first
Before any customisation, we build an evaluation set from your own examples. Every option is scored on it: the base model, the tuned instructions, retrieval and the fine-tuned model. You see the scores, the failures and the cost before you choose.
03 — Your data
04 — Serving and customised solutions
Served by OneVeer
A fine-tuned or distilled model is packaged as GGUF so OneVeer can serve it on your CPU or GPU behind a published alias. Applications keep calling the alias; returning to the previous model is a configuration change, not a release.
Onega products
OneVeer policies, aliases and guardrails; OneMail categories, routing rules, templates and connectors; OneDesk model providers. Each configuration is tested against your cases before go-live.
Solutions you already own
AI features in your CRM, service desk or document platform, tuned and evaluated on your cases and, where they allow it, routed through OneVeer for one policy and one audit trail.
05 — How it runs
An evaluation set from your examples, baseline scores for candidate models and methods, a data and licence review, a recommendation and a fixed-price proposal.
Instructions, retrieval or training as recommended, in your environment, scored on the holdout sample.
The customised model on real work with people reviewing its output, measured against the baseline.
Model card, data record, evaluation results, regression suite, serving configuration and runbook.
06 — FAQ
Often not. Many tasks are solved by clear instructions, examples and retrieval. We recommend fine-tuning when the evaluation set shows a gap those cannot close, or when a smaller tuned model meets the same bar at lower cost on your own hardware.
Open-weight models whose licences permit your intended use, in sizes that fit your hardware. Hosted models can be customised where the provider offers it and your data policy allows. The licence and its conditions are stated in the assessment.
It depends on the task. The Model Evaluation Sprint tells you whether your examples are enough, and of good enough quality, before you pay for any training.
Ownership of weights trained on your data is agreed in the statement of work before training starts. The base model’s licence terms always apply as well.