On-Device Video AI · Savant Lite

Cross-Platform Video AI Mobile Apps for iPhone, Android and Desktop

You have a computer vision model, or a product to ship. We build and deliver the application on Savant Lite, our cross-platform framework: live video in from anywhere, computer vision running on the device's own NPU, and the annotated result shown on-screen or streamed out to remote viewers — across iOS, Android, macOS, and Windows.

Application delivery service · you bring the model, or we build it · you get all the source, licensed to ship in your product.

iOS · Android · macOS · Windows · On-device NPU inference · Live in · on-screen or streamed out · Savant (open source)
The Problem

The Model Was the Easy Part

You want CV on live video, on the device itself — real-time, private, working without a connection. The real build is everything around the model, and it's per platform, in hardware.

Get the video in

  • SDR & HDR capture
  • Live, RTMP & file decode
  • Hardware decode, without stalling

Run AI on it

  • Your model on each OS's NPU
  • Overlay compositing
  • In-sync live playback

Keep, review, send

  • Battery-efficient encode & record
  • Instant storage and frame-accurate seek
  • Clip export, or RTMP / WebRTC restream

Four operating systems. Four accelerator stacks. A different API for every one of those steps — a multi-team, multi-quarter build for the parts of the app users never think about.

What We Deliver

You Ship the App — We Build the Machine Under It

It's a delivery service, not a toolkit — the difference is who does the multi-quarter build.

What it isn't

A toolkit you assemble

Not an SDK you have to learn, and not a demo you have to productionize — both leave the multi-team, multi-quarter build sitting on your roadmap.

What you get

A finished, shippable app

We build and hand you the application on Savant Lite, our cross-platform framework, on all four platforms. You bring your CV model — or we develop it with you — and we deliver the app around it, end to end: source included, licensed to ship.

You leave with every line of source — framework and custom — and a license to use it in your product without limitations.

Framework source

Savant Lite, the cross-platform runtime the app is built on — yours to read, build, and keep.

Custom source

Everything we write for your product — the application code and the integrations around it.

A product license, no limits

Use the framework and the custom code in your product without limitations.

The Complete Pipeline

Video In From Anywhere — AI on the Device — Results Where You Need Them

One path, already solved on every platform — from the camera to the screen, or to remote viewers when you need them.

LLM in the Loop

When It Has to Think, Not Just See

Perception runs on the phone. The LLM is fed structured metadata — signals, insights, and actions extracted from the video — not the video itself. From a guided menu to a personalized assistant.

  1. 01

    Perception on-device

    Camera and models run locally. Video stays on the phone.

  2. 02

    Structured metadata

    Perception is reduced to signals, insights, and actions. The LLM never sees the video.

  3. 03

    Cloud LLM reacts

    A hosted model reasons over that metadata and picks the next step.

  4. 04

    In-app response

    Back on the phone: a prompt, a menu, or a full assistant.

Simple

Guided menu

Detected state drives a deterministic flow — the right prompt surfaces the moment the camera sees the trigger.

Advanced

Personalized assistant

A cloud assistant reacts to perception in real time and adapts to the user — coaching, guidance, or the next-best action.

Why It Matters

On-Device, and on Every Platform — You Don't Have to Choose

Three payoffs that turn the engineering into product advantage.

Private, instant, offline, free per inference

The video never leaves the device; there's no network latency, no connection required, and no per-frame cloud bill. For inspection in the field or feedback on the court, that isn't a nice-to-have — it's the product.

Skip the four-platform team

iOS, Android, macOS, and Windows from one codebase, with the per-OS and per-chip optimization already done underneath. No hiring and managing a mobile + on-device-CV + video-pipeline team to ship a single app.

Local, with reach when you want it

The result can stay right on the device's screen — or, when you need it, reach a coach, a control room, or an audience live. Most approaches make you trade on-device privacy and latency for remote reach; here you don't have to.

Who Builds It

The People Building Your App Build the Infrastructure It Runs On

Framework authorship you can read — the same discipline, on the phone and the desktop.

Same Team, Same Discipline — on the Phone Instead of the Data Center

Savant Lite comes from the team that authors Savant, the open-source framework for real-time video analytics. Our work on Savant is open for anyone to read, so you can judge the engineering before you commit.

Tauri · Rust core · ORT / LiteRT · native capture, codecs & NPU

Code You Can Read

We author the open-source Savant framework — judge the engineering discipline before you commit.

Not a Lab Demo

We ship real-time video systems that run unattended — the same production discipline, on the phone.

On-Device Is the Core Discipline

A Tauri app with a Rust core, models served through ORT and LiteRT, native capture, codecs, and NPU acceleration — not a side project.

Real-Time Video Is Our Day Job

The same people and the same practices that build Savant for the data center, applied to the phone and the desktop.

Fit Check

Who It's For — and Who It Isn't

Sorted by one question: does it have to run on-device?

This is for you if

  • You're a B2B, enterprise, or funded-startup team building a real-time video-AI product that has to run on-device — because it needs to be real-time, private, offline-capable, or free of per-inference cost.
  • You don't want to build and manage a four-platform mobile + on-device-CV + video team to ship it.

This probably isn't a fit if

  • Your computer vision runs fine in the cloud, in batch, with no need for on-device real-time, privacy, or offline operation.
Lead use case

Sports & fitness

Technique and form analysis, player and object tracking, live stats and overlays, instant on-court feedback, and broadcast with AI graphics.

Lead use case

Mobile inspection & field service

Live defect detection on a phone or tablet, evidence recording, offline in the field, fast review and export, and a remote expert watching the annotated stream live.

Two Ways In

Bring Your Model, or Start From the Idea

Either way, you get a shipped, on-device app at the end.

You have models

We build the app around them

Your models, served on-device across all four platforms, inside a finished application we deliver end to end.

You don't yet

We develop the models too

And if you're earlier than that, we can de-risk the computer vision first, then harden the pipeline — before we ship the app.

This page is where the ladder ends — a shipped, on-device app. Earlier steps: Feasibility Sprint (does it work?) → PoC → Production (make it run) → this.

How It Works

From First Call to Shipped App

The call is where you see the framework live — then we scope it and deliver.

  1. 1

    Call + walkthrough

    We walk you through Savant Lite live, and learn your product, platforms, and constraints.

  2. 2

    Scope it together

    Platforms, inputs, models, and distribution — we define the app and how it ships.

  3. 3

    We deliver, end to end

    The finished app, all source — framework and custom — and a license to use it in your product without limitations.

Pricing is scoped to your app — platforms, models, and distribution. We confirm it before you commit.

Let's Build Your App

Book a call and we'll scope it — platforms, models, distribution — then deliver the app, the source, and a license to ship it in your product.

Not sure it's feasible yet? Start with a Feasibility Sprint, then PoC → Production.