Back to projects
In development · 2026
Third Eye AI
A voice-first AI built into the Third Eye mobile app and trained for blind users. Strips the visual and color-based assumptions out of mainstream AI and rebuilds understanding around what matters: sound, space, environment, intent. First-user release in 2026.
What it is
What is it?
Most consumer AI is trained on a sighted person's questions about a sighted person's world. "What does this look like?" "Find the red one." "Show me the chart." The default mental model is visual.
Third Eye AI starts somewhere else. It is trained on the questions blind users actually ask — about navigating an unfamiliar room, identifying a sound, parsing a conversation, reading a label, finding a friend in a crowd, knowing what's in front of you and what's behind it. Voice in, voice out. No assumption that the user can confirm anything by glancing at a screen.
Third Eye AI starts somewhere else. It is trained on the questions blind users actually ask — about navigating an unfamiliar room, identifying a sound, parsing a conversation, reading a label, finding a friend in a crowd, knowing what's in front of you and what's behind it. Voice in, voice out. No assumption that the user can confirm anything by glancing at a screen.
How it works
How does it work?
Third Eye AI lives inside the Third Eye mobile app. Users speak; the model listens. Output is spoken back, with optional haptic confirmation for environments where audio isn't safe or appropriate.
Underneath, the model is fine-tuned on a corpus of blind-user interactions, voice queries, and environmental audio — building up a different intuition about what a "useful answer" looks like. We're collecting that training corpus in partnership with the community, with full consent and full transparency about what's used and what isn't.
The app is designed for the same low-bandwidth conditions as the rest of our work. Where the cloud isn't reachable, a smaller on-device model still answers the most common queries.
Underneath, the model is fine-tuned on a corpus of blind-user interactions, voice queries, and environmental audio — building up a different intuition about what a "useful answer" looks like. We're collecting that training corpus in partnership with the community, with full consent and full transparency about what's used and what isn't.
The app is designed for the same low-bandwidth conditions as the rest of our work. Where the cloud isn't reachable, a smaller on-device model still answers the most common queries.
The reason
Why we built it
The blind users we work with already use mainstream AI assistants. They also describe the same frustrations over and over: assistants that default to visual descriptions, refuse to commit to spatial answers, and produce a kind of confident vagueness that's harder to use than silence.
Third Eye AI is built around a single design principle: a useful answer to a blind user is rarely the same as a useful answer to a sighted user. Same model architecture. Different training. Different defaults. Different definition of "helpful."
Third Eye AI is built around a single design principle: a useful answer to a blind user is rarely the same as a useful answer to a sighted user. Same model architecture. Different training. Different defaults. Different definition of "helpful."
In practice
How is it used?
The app is in active development for first-user release in 2026. Once shipped, users will be able to ask the model anything — practical questions ("what's on this label?"), spatial questions ("what's the room I just walked into?"), social questions ("who's at this table?"), and abstract questions, the same way any user would talk to any other AI. The app integrates with the Third Eye Kit hardware where available, and works standalone on any phone where it isn't.
What’s next
What’s the future of it?
A small private beta with users from our existing cohort, then phased rollout in 2026. The training corpus is being built now. If you're a blind or low-vision user who'd like to contribute voice queries (paid, with full consent and full data control), we'd like to hear from you.
Support projects like this.
Every dollar funds open-source tools that are free at point of use — and stay that way.