Inspiration
Growing up in Uzbekistan, I noticed a massive communication gap facing the Deaf and hard-of-hearing community. Most people do not know Uzbek Sign Language (UZSL), which creates an invisible barrier for deaf individuals trying to access education, navigate daily life, or share their voices. After discovering the International Olympiad in Artificial Intelligence (IOAI) and seeing how advanced machine learning can solve real-world problems, I was inspired to create SADO (meaning "Voice" or "Echo" in Uzbek). My goal is to build an open-source tool that turns silent hand gestures into an active, real-time voice.
What it does
SADO is a real-time computer vision pipeline designed to bridge the gap between Uzbek Sign Language and spoken or written language. Using a standard device webcam, the system tracks hand movements, analyzes complex gesture coordinates, and instantly translates those sequences into text.
How I built it
The architecture of SADO is built to run entirely on accessible, consumer-grade hardware. The development stack includes:
- Frontend: HTML5, CSS3, and JavaScript to manage a fast, clean user interface and handle the webcam stream feed.
- Hand Tracking: Google's MediaPipe framework to extract multidimensional coordinate arrays (tensors) from hand landmarks in real time.
- Backend Processing: Python and NumPy to clean, scale, and shape the data matrices.
- Deep Learning Engine: A custom-designed neural network architecture built from scratch using pure PyTorch to handle sequence classification and map coordinates to exact UZSL definitions.
Challenges I ran into
One of the biggest hurdles is that Uzbek Sign Language is a low-resource language dialect in the field of AI, meaning there are no massive pre-made datasets available on the internet. Additionally, ensuring that the machine learning model remains lightweight enough to calculate data frames instantly on a standard laptop without massive GPU costs requires intense optimization of tensor shapes and model layers.
Accomplishments that I'm proud of
- Teaching myself advanced machine learning frameworks like PyTorch and NumPy at 16 years old.
- Transforming a personal failure at a national science competition into the fuel needed to independently research and engineer a highly sophisticated medical/accessibility application.
- Designing a clean, modular pipeline where a JavaScript frontend can seamlessly communicate data to a heavy Python deep learning engine.
What I learned
I deeply learned that great AI isn't just about importing a library and writing a basic loop—it requires a profound understanding of mathematical definitions like backpropagation, data normalization, and matrix dimensions. I also learned how to structure a large-scale project portfolio using Devpost and GitHub to mimic professional engineering standards.
What's next for SADO
SADO is currently in its active alpha development stage! My immediate roadmap includes:
- Completing the full Scrimba-based JavaScript frontend interface layout.
- Building out the custom dataset pipeline for key Uzbek Sign Language gestures.
- Training and fine-tuning the PyTorch classification engine.
- Showcasing the fully operational build at the next national selection stages for major tech initiatives and international scientific olympiads.
Log in or sign up for Devpost to join the conversation.