Deliver V1 before expanding the demo
Build an Android Decision Lab project with reproducible setup, a fixture mode, a backend integration and a Compose UI. Our companion implementation is planned for the Android Engineers Android AI Cookbook; the Jev sample is not yet available. V1 includes voice or text input, synthetic notification triage, action previews and the Decision Inspector. Voice should have a text fallback; live notification access is optional. CameraX belongs to V2 and AppFunctions to V3.
Document which parts you implemented and which remain planned. If provider access is unavailable, the fixture implementation is still useful architecture work, but it is not a completed live Jev integration. Keep that distinction in the README and recording.
Acceptance scenarios
| Scenario | Required visible outcome |
|---|---|
| Supported note request | Reviewed transcript → route → editable draft; no automatic save |
| Ambiguous request | Clarification; no invented arguments |
| Notification with read-aloud disabled | No speech even if the model recommends it |
| Headphones disconnect during inference | Execution policy is checked against current state |
| Invalid or timed-out response | Explicit error and preserved input; no action |
| New request supersedes old one | Old result is recorded but cannot execute |
| Duplicate result or repeated callback | One execution at most for the action identifier |
| Denied microphone or camera permission | Usable text flow and accurate availability state |
Run these with deterministic fixtures first. Then exercise live inference with fictional data and keep its evidence separate. Do not test by sending another person's private notifications to the backend.
Package the evidence
Include setup instructions for Android and the backend, example configuration without secrets, device requirements, provider setup links and the selected model version. Explain your decision schema, permission gates and failure states. Record any measured limitations instead of assuming advertised performance applies to your app.
Save a labeled evaluation dataset and redacted results. Provide the sample count, routing accuracy, clarification rate, incorrect executions, failure rate and latency distribution. Describe how you tuned thresholds and which cases were held out. Keep the small rules baseline so reviewers can judge whether inference adds value for this scope.
Record a 30–60 second demo: show a successful decision, open the inspector, then demonstrate a blocked or ambiguous case. Display actual run timings. Explain the feature in one sentence: “Android provides context, Jev proposes a bounded decision, and Kotlin validates and executes the action.”
Score your readiness
Score each dimension 0, 1 or 2: absent, partially demonstrated, or reproducible with evidence.
- Architecture: fixtures and live responses share a domain boundary independent of Compose.
- Integration: a real redacted provider response maps correctly into the application contract.
- Execution: permission, argument, freshness and duplicate checks prevent unwanted actions.
- Evaluation: labeled cases and calculated results support the claims in the README.
- Communication: the demo accurately separates simulated data, model recommendations and completed actions.
A suggested portfolio target is 8/10 with full marks for execution. This is a self-assessment rubric, not certification. Missing live access remains a named integration gap even if the fixture-based UI is polished. Repeat failed acceptance scenarios before adding features.
Your next milestones
Publish the source and an accompanying build article when your own checks pass. Use the same architecture diagram and reproducible evidence in a codelab or talk. Add V2 perception only after V1 failures are understood, then attempt V3 AppFunctions with a documented supported environment.
Once our Jev sample is published in Android AI Cookbook, link its tested revision and setup instructions in the demo and accompanying article. Until then, describe the companion sample as planned and link your own implementation evidence for the assessment.