AI Bidet

As part of a consulting contract with SmartHygiene Inc., a Boston-based startup, I tackled a unique challenge: developing an AI-powered bidet system that respects user privacy while delivering advanced functionality.

Privacy concerns made traditional data collection unfeasible. To navigate this, I generated a synthetic dataset using 3D character models in Blender (bpy). By annotating specific body parts, we created a comprehensive training set without compromising real user data.

Performance was paramount. I implemented extensive data augmentations to enhance the model's accuracy and robustness. Introducing a noise injection scheme stripped texture information from the synthetic images, allowing us to merge them effectively with a limited real-world dataset.

To streamline our workflow, I set up a CVAT server for our team of annotators. This facilitated efficient labeling with precision.

Deploying the AI on an embedded device demanded optimization. I quantized the models for inference using TensorFlow Lite and the Lite Runtime, ensuring seamless operation on hardware with limited resources.

This project was more than just coding—it was about innovating within constraints. We addressed privacy head-on, engineered solutions where none existed, and brought a real AI product to life. Working on the AI Bidet was a fulfilling experience that pushed the boundaries of AI and embedded systems, turning challenges into achievements.