Saha prototype presents local, voice-guided yoga coaching with stated privacy limits

Current assessment: The available record supports describing Saha as a functioning, open-source prototype for webcam-based yoga guidance that is designed to process camera input locally and speak pose cues aloud. It does not support describing the project as independently validated, broadly tested, or ready to make reliable alignment or safety judgments across users and settings.

Reported facts

Saha was published on Hackster.io on Aug. 17, 2026, by Darcy VanPelt. The project is presented as a Java 26 desktop application for an approximately 20-minute beginner yoga practice. According to the project page, it uses MoveNet SinglePose Lightning to estimate body landmarks from a user’s webcam, overlays landmarks on a mirrored preview, and provides spoken instructions and limited corrective cues.

The creator states that normal operation does not save video, images, facial data or landmark coordinates, and that only derived session metrics are stored locally in a JSON file. The same account says the application has no account requirement, cloud upload path or API key requirement. A public GitHub repository linked from the project page contains source code, installation instructions, technical documentation and a stated MIT license.

The repository further describes a confidence-gating approach: when camera input is unreliable, the application is intended to pause or provide instruction-only guidance rather than present an alignment judgment. It also documents a demonstration fallback using synthetic landmarks when the model is not installed.

Supporting evidence

The strongest evidence is the public source repository, which documents the application’s stated architecture, local persistence approach, webcam workflow, pose routine and testing setup. The repository also contains a detailed limitations section rather than presenting the system as comprehensively proven. That disclosure is material because it narrows what the project’s technical claims establish: the codebase and design documentation support that a prototype was built and publicly released, while the implementation claims remain primarily creator-reported.

The Hackster.io project page independently mirrors the core description and identifies the project as a Java-based desktop coach. It states that the system uses local pose estimation, operating-system speech synthesis and locally stored derived metrics. The page also says the project was published under an MIT license and links readers to the public repository.

Limitations and competing evidence

The most important limitation is disclosed by the creator: camera coaching was tested against one person, on one Windows machine, in one room, with one camera height, one lighting environment and one set of body proportions. The repository expressly says this was not a validation study. It identifies floor poses as the weakest scenario because the underlying pose model is described as being trained on upright people, and it says single-view, two-dimensional angles cannot establish depth, guarantee alignment or guarantee safety.

There is also no independent assessment in the consulted record from yoga teachers, accessibility specialists, privacy auditors, clinicians or a research institution. The project’s statements that no data is uploaded and that only derived metrics are retained are credible descriptions of the stated design and publicly available code, but they are not equivalent to an external privacy audit. Likewise, a listed automated test suite demonstrates development effort, not real-world coaching accuracy across varied bodies, camera placements, physical abilities or home environments.

No location was identified in the available materials. The available record also does not establish whether Saha has users beyond its creator, has been released as a packaged consumer application, or has received independent technical review.

Current conclusion

A cautious reading is that Saha is a real, publicly documented prototype exploring a meaningful design choice for digital yoga instruction: combining local computer-vision processing with spoken, confidence-limited guidance rather than cloud-dependent video analysis. The available evidence supports the narrower claim that its creator has released code and documentation for that approach. It does not establish that the system delivers dependable pose correction, improves practice outcomes, works equitably across users, or provides verified privacy protection. Independent trials involving varied participants and settings, review by qualified yoga instructors, and a technical privacy or security assessment would materially change the current assessment.