Aevum Flow, User Journey, and System Architecture

This page captures three things in one place: current data flow, a realistic user journey, and the repo-supported system architecture including the documented Cloudflare R2 model-download path and edge-device execution model.

Current code evidence Repo-documented architecture Gap / wiring risk Edge-device / download path

1. Real Flow Diagram

User Input Voice / text / widget / shortcut Dashboard Path Current strongest baseline path Capture / Import Path Pipeline-dependent path Risk Pipeline not booted on normal start BehaviourIntelligenceLayer Dashboard route IntelligencePipeline Used by capture/import/shortcut flows InstantProcessor MemoryEngine EventQueueActor queue / validation / merge SwiftData / SQLite Local Store MemoryEntry, AevumEvent, persona, habit, legacy, graph state Local API 127.0.0.1:8080 Single prompt endpoint Visible Surfaces Dashboard / Timeline / Knowledge Map Widgets / App Intents

This is the real current-state flow. The dashboard path writes more directly and is safer than pipeline-dependent flows.

2. User Journey

Install / Open iPhone / iPad Onboarding Persona / privacy / setup Unlock / Enter Biometric gate if enabled Think or Capture Voice / text / capture sheet Store / Retrieve Timeline / graph / dashboard Optional model download path R2-hosted Phi model shards download locally Edge Device runs model locally Current journey gap Capture-sheet and shortcut direct logging rely on pipeline boot that is not in normal startup flow

This journey shows the actual iOS-first experience, plus the optional model-download branch evidenced by the repo’s local model manager and MLX architecture docs.

3. System Architecture

Edge Devices iPhone / iPad current Mac absent in app target App Shell Onboarding / privacy / biometric / tabs Dashboard / Timeline / Knowledge Map / Settings Local API 127.0.0.1:8080 Single prompt endpoint Cloudflare R2 Repo-evidenced model shard host phi-core shard source Behaviour / Intelligence Layer BehaviourIntelligenceLayer InstantProcessor / queue / daily processor Local Storage Layer SwiftData / SQLite / UserDefaults / local files AevumSecureGraph.store Model Runtime Layer Qwen bundled locally Phi shards downloaded then mounted locally Widgets / Shortcuts Quick Capture widget App Intents Integration Layer Google / Outlook / Notion Currently stub/mock Native Services Speech / biometrics / EventKit iOS-first

Cloudflare R2 is included because the repo and ML architecture docs explicitly reference an R2-hosted phi-core shard source and a local SLMDownloadManager that downloads those shards onto the device before local model load.

Architecture Notes

What is proven in code

  • Local SwiftData store
  • iOS app shell and widgets
  • Local API on 127.0.0.1:8080
  • Local model download manager
  • R2 base URL in download manager

What is documented and repo-supported

  • Bundled Qwen bootstrap layer
  • Downloaded Phi reasoning layer from R2
  • Edge-device execution model
  • Single-model-in-memory switching constraint

What is still a control risk

  • Pipeline-dependent flows are not wired as safely as the dashboard path
  • External integrations remain mock/stub level
  • Mac is still absent as a real app target