Wait for one answer
Each large target remains until an accepted response, so the user controls the pace.
Case study · Applied AI
SeeNA
A voice-guided iPhone prototype that helps a person complete qualitative vision tasks, with local scoring and a clear review of every answer.
- For
- People who need a simple, guided first step, including older adults and people who find written instructions difficult.
- My contribution
- I built SeeNA end to end as part of a four-person team for Syncs Hackathon 2026.
- Status
- Prototype
The brief
Make a first vision task easier to complete alone.
Small instructions and conventional eye charts can make a first vision task difficult to complete alone, particularly when sight or access to help is limited.
SeeNA guides positioning by voice, presents one target at a time, listens for an answer and lets the person review what was heard. The phone owns scoring; the language service has a bounded supporting role.
My role: I built SeeNA end to end as part of a four-person team for Syncs Hackathon 2026. Team: Karthik Ramesh, Kishore Srinivasan, Suryateja Challa and Sujan Ramesh.
What people do
One clear task at a time.
- 01Press Start
- 02Settle at the guided 40 cm distance
- 03Answer one Landolt C target at a time
- 04Complete the Gabor orientation task
- 05Review each eye's qualitative result
- 06Compare every answer with the recorded response
The decision that shaped it
The phone owns the result
Language models can transcribe bounded answers and explain allowed facts, but cannot create or change a score.
How I built it
The phone keeps hold of the result.
The architecture follows the product decision. Each part has a job people can inspect.
- Guidance
SwiftUI flow, spoken prompts and haptic feedback
- Readiness
TrueDepth distance and Core Motion quality checks
- Tasks
Pixel-rendered Landolt C and Gabor targets
- Scoring
Deterministic local scoring and answer evidence
- Language
Bounded transcription and checked qualitative explanation
- Review
Per-eye task outcomes and a complete answer audit
Further choices
The boundaries I kept explicit.
Show what was heard
Answer review exposes recognition mistakes instead of hiding them inside a final result.
Evidence and limits
What the prototype can and cannot show.
Built in
SwiftUI · ARKit · Core Motion · Swift · TypeScript · OpenAI
Product safeguards
- Local task scoring
- Distance and session-quality gates
- Bounded language output
- Per-answer review
Available evidence
- Native SwiftUI prototype
- Voice-guided task journey
- Local scoring with answer review
Current boundary
- This is a non-diagnostic, qualitative prototype. It does not measure clinical acuity or contrast sensitivity, prescribe lenses or rule out an eye condition.
- Physical sensing needs a compatible Face ID iPhone. Simulator and deterministic checks do not establish clinical validation.
The product in use
Every recorded answer stays reviewable.
