Velora

Smartness v2 — the self-learning release

Spoken self-corrections, done right. "Let's meet at 3pm — no no, 6pm" → Let's meet at 6 p.m. Works for any cue — "oh sorry, Shubham", "I mean Rahul", "scratch that", "delete this line" — across sentence boundaries, driven by the LLM's understanding of speech repair (a replacement swaps a parallel item: a name for a name, a time for a time). Emphasis ("no no no!"), real apologies, and content "no"s are preserved.

Real-time learning loop (all local). Edit the text Velora just pasted and it learns your fix within seconds — the HUD pill returns with the mishearing struck out → your correction → Learned. Misheard names commit on the first fix; everything is manageable (and deletable) in Settings.

It mines your vocabulary while idle. When nothing is happening, the local LLM extracts names and jargon from your dictation history into an auto-vocabulary that biases recognition itself (whisper initial_prompt) — "Velora" stops coming out as "valora" at the STT level.

Flat latency at any length. Long dictations are decoded and cleaned in pause-aligned segments while you speak: a 75-second dictation finishes ~1.5s after you stop (previously fell back to raw, unpunctuated text) — with live partial transcript in the HUD, now also for whisper.

Terminals got smart. Long prose dictated into a terminal (Claude Code, codex chats) is cleaned and punctuated; short shell commands stay verbatim.

Everything runs on-device. Zero network. Zero telemetry.

180 engine tests; adversarially reviewed (2 independent models, 8 findings fixed); real-model eval 15/16 on the self-correction suite.