flowy is an AI-powered yoga instructor that enhances your yoga practice by analyzing your poses in real time and providing personalized suggestions for improvement. For each pose, it assigns a score based on the accuracy of your execution, using three precision levels (10, 50 or 100) which represent the percentage of pose accuracy.
'Things That Think': prototype an interaction leveraging large language models (LLMs), agents, or machine learning techniques. This entailed exploring creative methods for utilizing inputs such as cameras and microphones to detect events, body positions, speech, sounds, and more. These inputs could then drive various forms of output, which may included web interfaces, images, sound, text, speech or even physical interactions
An object designed to assist individuals practicing yoga at home while offering real-time feedback. We envisioned a compact device that could function either as a smartphone accessory (app + phone stand) or as a standalone unit with its own screen. In both cases, portability was key, ensuring users could enjoy yoga sessions wherever they are.
The primary input is the front-facing camera, which tracks the user movements. The outputs include both visual elements (such as a timer, pose illustration and performance score) and auditory feedback to guide the user in real-time.
The interface is designed to be as simple and minimal as possible, ensuring that key information remains clear and easily visible from a distance of approximately 2.5 meters, where the user will typically be positioned.
It displays three crucial pieces of information:
The illustrations of the poses were drawn from scratch, featuring round, playful shapes to better align with the overall branding. The animation of the pose that appears on the final screen was created using Adobe After Effects.
Tree Pose
Crescent Lunge
Pose Animation
We explored different technologies that could support our goal. After experimenting with different tools and ML models (e.g., Teachable Machine, LLAVA), we discovered that GPT-4o was effective enough for our needs.
During this phase, we also considered integrating skeleton detection with MediaPipe to enhance feedback accuracy. However, GPT-4o proved highly reliable in correctly identifying yoga poses.
Contains two key functions that use OpenAI GPT-4 API:
The 3D model was created in Fusion 360, featuring a custom-designed structure to securely hold the phone, and was subsequently 3D printed.
The shape of the model was inspired by the starting yoga pose, with the crossed legs serving as a support base and the hands joined above the head acting as a handle, also making the object more manageable and portable.
Model Sketches
Rendered Model
Brand Identity