Building Varmir in public

Published the documentation — quickstart, full API reference and client guides (JavaScript, Python, cURL) — plus an in-browser live demo and the marketing site. Added ⌘K search across the docs.

A WebSocket endpoint that returns partial transcripts in ~200ms as you speak, sends audio as compact binary frames, and runs a composed pipeline: transcription → optional translation → optional prompt shaping — each stage toggled independently.

A lightweight, torch-free runtime built on CTranslate2 (INT8) and llama.cpp. GPU by default with automatic CPU fallback, packaged as a single universal Docker image — the same engine scales from a laptop to a server.

Language routing pairs a fine-tuned model for Ukrainian with a broad multilingual base, and NLLB-200 adds translation across ~96 languages. Dedicated language detection keeps closely-related languages (like Ukrainian and Russian) from being confused.

Fine-tuned our speech model on Common Voice and FLEURS with data augmentation, reaching roughly 11.6% word-level error rate on the Ukrainian test set — with punctuation included out of the box.

Started building a real-time, multilingual speech-to-text engine with built-in translation — accurate, fast, and simple to integrate through a developer API.