Skip to main content

5-layer pipeline deep dive

The Aucert platform processes mobile application testing through a 5-layer AI pipeline. Each layer is a distinct processing stage connected by the MCP (Model Context Protocol) message envelope.

MCP message envelope​

All inter-layer communication uses the MCPMessage protobuf (defined in proto/pipeline.proto):

message MCPMessage {
  string task_id = 1;
  string source_layer = 2;
  string target_layer = 3;
  bytes payload = 4;
  map<string, string> context_snapshot = 5;
  double confidence_score = 6;
  string trace_id = 7;
  int64 timestamp_ms = 8;
}

Layer 1: Generation​

Location: backend/platform/src/main/kotlin/.../domain/generation/

The Generation layer takes input from the Knowledge Graph and produces test scenarios. It analyzes app structure (screens, navigation flows, API contracts) and generates comprehensive test plans that cover:

  • Critical user paths
  • Edge cases from node property combinations
  • Regression scenarios from historical bug patterns
  • Accessibility checks

MVP model: Claude Sonnet (single model, no routing yet).

Layer 2: Execution​

Location: backend/platform/src/main/kotlin/.../domain/execution/

The Execution layer runs generated test scenarios on Android emulators. It performs:

  • App installation and launch
  • UI interactions (tap, swipe, text input, scroll)
  • Screenshot capture at each step
  • Execution trace recording with timing data

Device Twin overlay: Planned for Phase 2. Will augment emulator results with predicted real-device behavior.

Layer 3: Analysis​

Location: backend/platform/src/main/kotlin/.../domain/analysis/

The Analysis layer applies visual reasoning to screenshots and execution logs. It compares expected vs actual behavior using multimodal AI models, detecting:

  • UI regressions (layout shifts, missing elements, color changes)
  • Functional failures (wrong screen, error states, crashes)
  • Performance issues (slow transitions, unresponsive UI)

Layer 4: Decision​

MVP: Confidence threshold only — if the Analysis layer's confidence score exceeds the configured threshold, the test passes. Below threshold, it's flagged for review.

Full version (Phase 2+): 4-stage Verification Cascade with escalating cost and rigor. See Verification Cascade.

Layer 5: Reporting​

MVP: Dashboard page showing test results, bug reports, and metrics.

Full version (Phase 2+): Structured bug reports with Jira/Linear integration, team notifications, and feedback loops into the Knowledge Graph.

Interface + adapter pattern​

All layers communicate through interfaces, not concrete implementations. This enables:

  • Local mode (monolith): All layers in one JVM process, direct function calls
  • Remote mode (microservices): Each layer as a separate service, gRPC calls

A config flag (aucert.pipeline.mode=local|remote) switches between modes. See AGENTS.md for interface rules.

What's next​