Oculi·Dei
Oculi Dei LLC — Michigan

Adapted models

Private Evaluation

Turn your token audit into a working deployment plan. We evaluate vocabulary-adapted inference for structured extraction and tool-call workloads: the baseline and the adapted model on the same held-out tasks, measuring useful results and serving costs, to find where a private deployment earns its keep.

What the service combines

Your application remains in control of business decisions and tool execution.

Schema-constrained generationTool calls are generated against the schema your request declares, so their structure matches it.
Exact source-copy rules, where you want themValues you designate are copied exactly from your source text rather than rewritten.
A bounded calculator for explicit arithmeticExplicit arithmetic is handled by a separately labelled calculator, not by the model.
You keep the decisionsThe model proposes; your application decides and runs every tool itself.

Start with your workload

Agreed in writing before anything is measured.

We agree on the task, the acceptance criteria and the deployment target, then measure quality and overhead-inclusive savings on held-out tasks before any production commitment. A token-count reduction alone does not establish a saving: the evaluation measures the whole request.

Measured latency for a model adapted to its lexicon is on the Lexicon Build page, with the conditions it was measured under. Your own workload is measured on your own target before anything is committed.

Where it stops

The limits are here rather than in a footnote, because finding them later is worse for both of us.

The audit figures describe vocabulariesThe reduction ratios published for the Token Audit are measurements of vocabularies on text. None of them is a property of an adapted model, and none is quoted for one.
Private and single-customerEvaluations and deployments are private, with your acceptance required. Each deployment target is qualified as part of the engagement. There is no shared service and no uptime guarantee.

Available for private evaluations

Tell us what you run

The workload you want to improve, the model you run today, and your current infrastructure.

Discuss a private evaluation →