Open-model interfaces

Rust / Python
Local inference
Model adapters

Volume

See the reading application ↗

Local model integration · personal product

Open voice and reading models become controls someone can use: choose a reader, play a book and correct a line.

LOCAL · WINDOWS ALPHA

01 — THE READING ROOM

Actual app recordings

Model configuration

VolumeModel profiles from the application

Voice / instant

Kokoro

A CPU voice profile. The app can read aloud without reserving GPU memory for speech.

CPU · bf_emma · Apache-2.0

An inspection of shipped configuration, not a live inference demo.

Behind the work

Hardware profiles connect the voice and reading model to memory, available devices and listening needs. The model registry records download sizes and credits; the interface keeps a chosen override. Text preparation, audio and word timings live in the book bundle rather than inside a model session.

Praxis

Resolver & transport note ↗

Model transport & capability configuration · experimental

Native model APIs and compatible endpoints have different capabilities. The interface keeps those differences visible.

transport / adapter boundariestext

anthropic-native   → native messages
openai-responses   → native responses
openai-compat      → compatible endpoint

provider + model + policy
             ↓
     resolved manifest
             ↓
       recorded session

Architecture summary from the transport implementation, not a live model call.

Behind the work

The compatibility adapter can target a local server through its configured endpoint. Native adapters keep provider-specific controls. Capability records, user policy and endpoint configuration remain separate.

← Portfolio Discuss a project ↗