Transformation 02 · Human-centered AI
Designing AI around a real accessibility need
The model was one part of the product. The harder question was how to make environmental sound information useful without adding distraction.
The human need
Environmental sounds carry information, but not every sound deserves an alert.
The project explored driving support for Deaf and hard-of-hearing people by classifying meaningful environmental sounds and translating them into visual cues.
A useful product could not simply replace every sound with a notification. Latency, false alerts, attention, and visual hierarchy were core product constraints.
The product diagnosis
This was an attention-design problem as much as a classification problem.
The team needed to decide which information deserved a driver’s attention, how quickly it should arrive, and how the interface should communicate a category without becoming another distraction.
We connected the research review, model behavior, and mobile experience around one outcome: timely, glanceable, differentiated information.
Decisions and build
Make model behavior legible inside a restrained mobile experience.
We used short sound windows to balance responsiveness with enough signal for classification. I built the React Native mobile experience and connected it to the audio-classification pipeline.
The technical path used YAMNet/AudioSet-derived categories and a prototype service, with a future path toward on-device deployment.
Evidence and boundaries
Accuracy is evidence—but it is not the whole product.
The model reached 92% accuracy on the project evaluation setup, and the team produced a functioning end-to-end accessibility prototype.
That benchmark does not establish clinical effectiveness or real-world road safety. The next product questions concern alert relevance, usability, latency, and the cost of specific false positives.
Reflection
In high-stakes AI, usefulness, reliability, attention, oversight, and recourse have to be designed together.
- What I would do differently
- I would bring representative users into prototype evaluation earlier and define alert-specific false-positive costs before expanding the sound taxonomy.
- Framework connection
- Product Diagnosis Canvas · AI Trust Framework
Evidence boundary
Back to all transformations This case describes an academic project and its public technical approach. It does not claim clinical deployment or validated real-world safety outcomes.