aspirant200715/neuromesh

Purpose and runtime

NeuroMesh is an offline Android crisis-response mesh. Each Android 10+ phone senses possible emergencies, creates local assessments with on-device Gemma 4 inference or heuristic fallback, exchanges structured assessments with nearby phones, reaches consensus, and presents actionable alerts. It uses Kotlin, Coroutines/Flow, Hilt, Room, CameraX, MediaPipe LiteRT, and Android Nearby Connections; no cloud/server is required.

App bootstrap and Android entry points are app/src/main/java/com/neuromesh/crisis/NeuroMeshApplication.kt, app/src/main/java/com/neuromesh/crisis/presentation/MainActivity.kt, and app/src/main/AndroidManifest.xml. Critical alerts have a dedicated activity at app/src/main/java/com/neuromesh/crisis/presentation/CrisisAlertActivity.kt.

Presentation and orchestration

The UI follows MVVM. Activities and custom views render camera, mesh state, observations, and alerts; ViewModels own lifecycle-aware collection and state. MainViewModel is the control-plane coordinator for the detection loop, incoming mesh messages, consensus events, and active alert state.

Key paths:

• app/src/main/java/com/neuromesh/crisis/presentation/viewmodel/MainViewModel.kt

• app/src/main/java/com/neuromesh/crisis/presentation/viewmodel/MeshViewModel.kt

• app/src/main/java/com/neuromesh/crisis/presentation/viewmodel/CrisisViewModel.kt

Local sensing and observation

Infrastructure adapters isolate Android hardware. Camera, microphone, and environmental sensors produce textual/signal descriptors rather than raw media; this is a material boundary because the Gemma model is text-oriented. The observer combines those descriptors into an Observation, with heuristic detection available when model loading/inference cannot run.

Key paths:

• app/src/main/java/com/neuromesh/crisis/infrastructure/sensor/CameraSensorManager.kt

• app/src/main/java/com/neuromesh/crisis/infrastructure/sensor/AudioSensorManager.kt

• app/src/main/java/com/neuromesh/crisis/infrastructure/sensor/EnvironmentalSensorManager.kt

• app/src/main/java/com/neuromesh/crisis/domain/agent/ObserverAgent.kt

On-device multi-agent reasoning

The domain pipeline separates perception, assessment, and guidance: ObserverAgent yields observations; ReasonerAgent fuses recent local and received observations into a SituationAssessment plus explainable trace; ActionAgent produces a CrisisAlert. Use cases establish the principal workflow and prevent presentation from directly coordinating infrastructure.

Gemma integration is contained behind runner/loading/prompt/parser boundaries, allowing structured-output recovery and fallback behavior.

Key paths:

• app/src/main/java/com/neuromesh/crisis/domain/agent/ReasonerAgent.kt

• app/src/main/java/com/neuromesh/crisis/domain/agent/ActionAgent.kt

• app/src/main/java/com/neuromesh/crisis/domain/usecase/DetectCrisisUseCase.kt

• app/src/main/java/com/neuromesh/crisis/infrastructure/ml/Gemma4ModelRunner.kt

• app/src/main/java/com/neuromesh/crisis/infrastructure/ml/OutputParser.kt

Peer mesh transport

Nearby Connections is encapsulated so domain code remains SDK-agnostic. The network manager advertises/discovers peers, sends and receives serialized messages, and tracks peer connectivity. Structured observations/assessments—not images, audio, or raw LLM output—cross the network boundary. Transport is Bluetooth/Wi-Fi Direct P2P, using a Bluetooth-preferred Nearby strategy.

Key paths:

• app/src/main/java/com/neuromesh/crisis/infrastructure/network/NearbyConnectionsWrapper.kt

• app/src/main/java/com/neuromesh/crisis/infrastructure/network/MeshNetworkManager.kt

• app/src/main/java/com/neuromesh/crisis/infrastructure/network/MessageSerializer.kt

• app/src/main/java/com/neuromesh/crisis/data/repository/MeshRepository.kt

Gossip and consensus

Received assessments enter a TTL-bounded, deduplicated gossip flow and are evaluated over a time window. VotingStrategy applies confidence-weighted voting; ConsensusEngine emits reached/disputed/no-consensus events and builds a merged assessment. A reached result triggers alert generation, then UI presentation.

Key paths:

• app/src/main/java/com/neuromesh/crisis/domain/consensus/GossipProtocol.kt

• app/src/main/java/com/neuromesh/crisis/domain/consensus/ConsensusEngine.kt

• app/src/main/java/com/neuromesh/crisis/domain/consensus/VotingStrategy.kt

• app/src/main/java/com/neuromesh/crisis/domain/usecase/ShareObservationUseCase.kt

Persistence, models, and dependency boundaries

Typed domain/data models carry observations, assessments, traces, alerts, and peers. Room persists observations and alerts for local continuity; repositories shield callers from DAO/network details. Hilt modules compose runtime implementations. The intended boundary is Presentation → Domain, with infrastructure owning Android SDK, ML, network, and storage access.

Key paths:

• app/src/main/java/com/neuromesh/crisis/data/model/Observation.kt

• app/src/main/java/com/neuromesh/crisis/data/model/SituationAssessment.kt

• app/src/main/java/com/neuromesh/crisis/data/local/NeuroMeshDatabase.kt

• app/src/main/java/com/neuromesh/crisis/data/repository/ObservationRepository.kt

• app/src/main/java/com/neuromesh/crisis/di/AppModule.kt