UViLox Alpha
Frame-level landmark classifier for alphabets and digits. Used as the fallback path when a sign falls outside the word vocabulary.
Uvilox AI is built on three pieces: a recognition model for Indian Sign Language, a realtime agent that turns what it understands into speech, and an action layer that can actually place the call or send the message. All of it is in active development — this page explains how it fits together, honestly.
None of these are publicly available yet. We publish scope and status only — no accuracy claims until we can share a benchmark we stand behind.
Frame-level landmark classifier for alphabets and digits. Used as the fallback path when a sign falls outside the word vocabulary.
Sequence model over hand, pose and face landmarks. Trained on a small in-house Indian Sign Language set that we are still expanding.
The layer that turns a recognised sign sequence into a spoken sentence and a real action — placing a call, sending a message, raising an SOS.
Our Indian Sign Language fingerspelling model is the first UViLox model available publicly. Show a letter to your camera or upload a photo, and the model tells you which letter it sees. Letters only for now — words and phrases are still in development.
Frames are sent to the UViLox recognition service only when you press a button, and are not stored by this page.
Supported vocabulary: ISL fingerspelling — letters A–Z and digits 0–9. Hold one clear hand shape in frame, with good lighting and a plain background, for the most reliable reading.
A guided walkthrough of how the UViLox realtime agent is designed to work on LiveKit: audio and sign frames stream into a room, the model reads intent, and a tool call places the call, sends the message, or raises the alert. This walkthrough is scripted — the live agent is still in development.
Waiting for input…
no tool invoked yet
Simulated walkthrough for illustration. No call, message, or alert is sent from this page, and no microphone is accessed.
A plain-language walk through the stack we're building. Numbers and benchmarks will follow once the models are stable enough to publish them.
Instead of raw video, we work on normalised keypoints. It keeps inference light enough for mid-range Android phones and means we never need to ship someone's face to a server to understand a sign.
Signs are motion, not stills. A temporal model reads a rolling window of keypoints and emits a phrase, with the fingerspelling model handling anything outside the word vocabulary.
Recognised phrases become an agent turn inside a LiveKit room — the same room that carries the call audio — so the AI voice, live captions, and the person on the other end stay in sync.
An assistant that can dial numbers needs guardrails. Actions are previewed before they run, emergency flows are separated from everyday ones, and the user keeps control of contacts and permissions.