| name | sign-language-ml-detection |
| description | Diagnose and improve Amy's Echo web sign-language ML detection quality. Use when working on MediaPipe landmarks, MLP inference thresholds, confidence/margin behavior, detector telemetry, fixture diagnostics, or profile-aware recognition regressions. |
Sign Language ML Detection
Use this workflow for webapp gesture recognition and ML decision-path changes.
1) Map the touched detection path
Prioritize these components:
- runtime detector and processing:
webapp/src/gesture/core/, webapp/src/hooks/useSignLanguageDetector.ts
- model load/inference:
webapp/src/gesture/installMlp.ts, webapp/src/gesture/modelClient.ts
- diagnostics and fixtures:
webapp/src/gesture/testing/fixtures/, npm --prefix webapp run diagnose:fixtures
- training-to-runtime contracts: shared frame/label types and server model version endpoints
2) Reproduce with deterministic fixtures first
- Run focused tests for the changed detection module.
- Run
npm --prefix webapp run diagnose:fixtures to generate comparison evidence.
- Validate confidence and margin behavior with fixture outputs before manual camera runs.
Use the checklist in references/detection-debug-checklist.md.
3) Protect profile-aware behavior
Always verify account/session and active profile context are both valid when detection depends on profile-trained labels.
Do not report success if:
- profile-specific model failed to load,
- server rejected auth/profile-scoped requests,
- fallback path masked regression without telemetry.
4) Validate user-visible reliability
- Keep caregiver/child-facing copy in German.
- Prefer abstention on low confidence over forced wrong class.
- Confirm diagnostics explain why a trained sign was not accepted.
5) Report ML evidence clearly
Include:
- threshold/margin values used,
- top1/top2/margin effects,
- fixture report path,
- tests executed and outcomes.
References
- Debug checklist:
references/detection-debug-checklist.md
- Detection test matrix:
references/detection-test-matrix.md