Pilot objectives
| Interaction | Target |
|---|---|
| Call join to two-way audio | P95 <3 seconds under supported network conditions |
| Call join to usable video | P95 <5 seconds under supported network conditions |
| Microphone/camera state feedback | <150 ms local UI response |
| Interim STT appearance | P95 <700 ms where network/provider permit |
| Simple response first audio | P95 <2.5 seconds after finalized turn |
| Camera-guidance acknowledgement | P95 <1.5 seconds after accepted frame event |
| TTS stop on interruption | P95 <300 ms |
| Reconnect after brief loss | Automatic, preserving session/case state |
| Persistence after accepted step | <5 seconds |
Media quality and adaptation
- Network adaptation may lower resolution, frame rate or video priority before ending the call.
- The interface always displays whether the AI currently has usable visual evidence.
- Selected diagnostic frames meet configured focus, exposure, crop-presence and subject-size thresholds.
- Full-resolution capture is explicit and acknowledged.
- Raw video recording is off by default unless approved and visibly disclosed.
Reliability
- No acknowledged turn or accepted frame is lost.
- Room and case state can recover independently.
- Provider timeout yields fallback, queue or explicit unavailability.
- Policy calculations are reproducible from stored inputs and rule version.
- At least 99% complete decision audits; 100% for treatment claims and selected-frame provenance.
Security and privacy
TLS/SRTP, encryption at rest, short-lived room credentials, least privilege, role/attribute controls, secrets management, media limits, prompt-injection defence, explicit recording state and auditable retention.
Accessibility
- Target WCAG 2.2 AA for documentation, farmer application and officer portal.
- Text and essential icons meet contrast requirements in light and dark themes.
- All controls have visible keyboard focus and screen-reader labels.
- Status is not conveyed by colour alone.
- Support text scaling, reduced motion, captions/transcript and touch targets suitable for low-cost devices.
- Urdu/English mixed-script content uses correct language and bidirectional isolation.
Maintainability
Provider-neutral interfaces, versioned schemas, automated validation, modular codebase and infrastructure additions only through measured need and an architecture decision record.
Usability
Farmer completion and comprehension are release criteria. Every uncertainty or degraded-media state exposes a clear next action. Long research persists after the call if necessary and returns an explicit status rather than an indefinite spinner.