Applied lab · AI, low-tech, hardware · Bidart, France

We build the tools that get people moving.

Movelytics is an applied lab with one goal: help more people move, every day. On-device AI when the problem needs it, low-tech when that works better, embedded hardware when the tool has to live in an object. When a camera is involved, no video leaves the device.

Movement is such great medicine.

Squat · knee, hip and ankle anglesIllustration. Authored pose sequence, not live inference.

  • 01iOS
  • 02Android
  • 03Web apps
  • 04Embedded hardware

01What we stand on

Four commitments. They hold for every product.

P1

On-device by default

Our models run on the user's phone, or on your own infrastructure. No video is uploaded for inference. No upload-and-wait analysis.

P2

Real time on the phones people own

We optimise for the devices that are actually in pockets, including low-end and five-year-old Android phones. Not only flagships.

P3

Your users' data stays yours

We meter usage for billing: a number of API calls. We do not store videos, keypoints or analysis history, and we never sell data.

P4

A lab that ships products

We train our own models and we ship them: SDK, SaaS, custom builds. Deeptech when the problem needs it, low-tech when that works better.

02Products

Three products, built and run by the lab.

Each one started from a problem we could not solve with what existed.

Flagship

PoseTracker

Movement understanding for fitness, wellness and rehab apps. It runs on the device and returns keypoints, joint angles, rep counts, a form score, jump height and flexibility measures. React Native, iOS, Android and web.

  • 1,250+ registered developers as of 17 November 2026.
  • In production in Flexifit, Kobi, How2Jump and MoveBreak.
  • Free tier of 200 API calls per month.
PoseTracker product video (silent)posetracker.com

Our own modelPrivate beta

PoseTracker Model

A real-time human pose estimation model that we train ourselves, designed for mobile processors. It ships first inside PoseTracker. A light version runs on low-end Android phones; a heavier, more precise one runs on recent phones.

39

keypoints

Body, feet, neck and shoulders, trunk front and back. Public models of the COCO family give 17.

18.4ms

per frame, model alone

Huawei FIG-LX1 2017 (Android 9), 4 threads

28.7fps

with the camera, over 10 minutes

Huawei FIG-LX1 2017, camera at 30 fps. Slowest 10 s window: 22.1.

Private beta; benchmarks are preliminary and measured on an internal build (17 November 2026).

Low-tech on purpose

MoveBreak

Noa is a wellbeing assistant inside Microsoft Teams and Slack. It prompts active breaks of one to three minutes during the working day. No app to install, no generative AI, predefined scenarios. The best tech is the one people actually use.

  • 2,000+ employees supported and 30,000+ wellbeing moments since early 2026 (internal data).
  • About 300 breaks: movement, posture, eyes, focus, team cohesion.
Chat10:50

Noa

A short break is in order. What do we start with?

  • Express active break · 1 minute, 1 movement
  • Long active break · 3 to 4 minutes
  • Anti-stress break · 1 to 2 minutes, breathing
  • Later
Express active break

Noa

Hold on tight, we are moving the neck.

Targeted zones

neck · shoulders · shoulder blades

Level

Medium

Timer

Noa

Break done.

Your next reminders stay planned at 13:40 and 15:50.

Looped illustration of the Noa scenario, accelerated. Neutral chat frame.See the full demo on move-break.com

03Cloud AI vs. Movelytics

Where the model runs changes everything else.

A comparison of two approaches, not of named companies.

Cloud AIMovelytics
Where inference runsOn a remote server, after an uploadOn the user's phone, or on your own infrastructure
LatencyA network round trip for every result; often upload, then waitPer frame, in real time: 28.7 frames per second processed with the camera on a low-end Android phone (10-minute average; slowest 10 s window 22.1; preliminary)
What leaves the deviceThe video, or frames taken from itNothing from the camera. Only usage metering: a number of API calls
Cost modelServer compute that grows with every minute of videoInference runs on hardware the user already owns; billing follows metered usage
Offline useNot possible without a connectionThe model needs no network to run
The frame rate is from our benchmark of 17 November 2026: Huawei FIG-LX1 2017, camera at 30 fps, 4 threads, internal build. Details on the benchmarks page.

05Proof

Numbers, with their dates.

1,250+

developers registered on PoseTracker

As of 17 November 2026

2,000+

employees supported by MoveBreak

Since early 2026, internal data

30,000+

wellbeing moments with Noa

Since early 2026, internal data

39

keypoints in the PoseTracker Model

Private beta, internal build

06Work with the lab

We also build it for you.

The team that trains our models takes on client projects: custom keypoints and fine-tuned pose models, domain language models that run on the device or on your premises, SDK and app builds, and hardware plus software products. Hardware client projects are under non-disclosure agreements, so we do not describe them here.

  1. 01

    Scoping call

    You describe the movement, the device and the constraint. We say what is feasible.

  2. 02

    Proof of concept

    A build you can run on the target device, with measured results.

  3. 03

    Production

    Integration, optimisation for the real device fleet, and maintenance.

07FAQ

Questions we are asked first.

What is Movelytics?

Movelytics is a French applied-AI lab based in Bidart, in the Basque Country. It builds real-time, on-device AI for human movement and ships it as products: PoseTracker, the PoseTracker Model and MoveBreak. It also builds custom models and products for companies in fitness, wellness, rehab and connected devices.

What does on-device AI mean at Movelytics?

The model runs on the user's phone or on the customer's own infrastructure. The camera image is analysed where it is captured and is not uploaded to a Movelytics server for inference. The technology page explains the full pipeline.

Does Movelytics store our users' videos or movement data?

No. Movelytics collects usage metering only: the number of API calls, used for billing. It does not store end users' videos, keypoints or analysis history, and it never sells data. The app that uses our tools keeps its users' data. See our data commitments.

Which products does Movelytics make?

Three. PoseTracker is a pose estimation SDK and API for fitness, wellness and rehab apps. The PoseTracker Model is our own real-time pose model with 39 keypoints, in private beta. MoveBreak is Noa, a wellbeing assistant inside Microsoft Teams and Slack.

Can Movelytics build a custom model or product for our company?

Yes. The lab takes on custom pose models and keypoints, domain language models that run on device or on premises, SDK and app builds, and hardware plus software products. Projects go from a scoping call to a proof of concept, then to production. See work with the lab.

Is the PoseTracker Model available today?

It is in private beta. Its benchmarks are preliminary and measured on an internal build, not on a shipped model. Teams that need real-time tracking with more keypoints, or custom ones, can request access to the beta.

Movement is such great medicine.

Tell us what your product should understand about movement.

A member of the team reads every message.