mozi-app.nexisfoundation.workers.dev
Aura AI
A personal, agentic AI assistant across iOS, Android and web — chat, rich interactive cards, durable automations, memory, connectors, voice, and an Aura-default model route through Vercel AI Gateway.
Aura AI
Agentic life copilot with governed workflow maps
07
Project context05
Problem
Personal AI assistants often become disconnected feature piles: chat, memory, connectors, automations, voice, and cards evolve separately. The hard part is making them one governed system with clear runtime boundaries.
Solution
Aura maps an agentic life copilot across clients, runtime, tool catalog, memory, connectors, voice, automations, and data layers so each capability has an explicit place in the architecture.
Challenges
The system has to keep context, connector permissions, voice, memory, durable automations, and rich cards synchronized across mobile and web without turning every feature into bespoke agent logic.
Innovation
The map emphasizes product architecture as much as model choice: it makes the assistant inspectable through services, layers, and run surfaces instead of presenting it as a monolithic chat app.
Domain expertise
This shows Stefan's product systems thinking across cross-platform AI assistants, Cloudflare/Supabase architecture, connector design, voice UX, memory systems, and durable automation flows.
Case study evidence11
Outcomes
- Turns a broad personal-assistant concept into an inspectable system map with clients, memory, connectors, automations, and voice in one model.
- Clarifies where product experience ends and runtime governance begins across web, mobile, and backend surfaces.
- Provides a strategic architecture reference for building an assistant that can remember, act, and automate safely.
Architecture decisions
- Cloudflare provides edge routing for chat, webhooks, and connector traffic.
- Supabase keeps identity, memory, relational state, and durable workflow records close together.
- Provider routing makes Aura the default picker path through Vercel AI Gateway while preserving GPT-5.5, hidden Anthropic fallbacks, Cohere rerank, and Apple-native surfaces.
Domain expertise signals
Personal AI OSMemory systemsConnector architectureVoice UXDurable automations
Technical deep dive09
Aura is a systems map for a personal AI operating layer. It shows how chat, cards, memory, voice, connectors, automations, and provider routing can live in one coherent product architecture.
Capability map
The architecture separates clients, runtime, tool catalog, memory, connectors, voice, automations, and data layers. That prevents the assistant from becoming a pile of unrelated features.
Memory and identity
Personal AI assistants need durable memory, user-scoped state, and identity-aware permissions. Supabase provides a relational foundation for continuity while still supporting explicit access boundaries.
Provider routing
Different models should not be treated as interchangeable. Aura's default picker id runs through Vercel AI Gateway to minimax/minimax-m3, while GPT-5.5, Anthropic fallbacks, Cohere rerank, and Apple-native capabilities remain routed by capability fit.
Product surface
Rich cards, voice capture, automations, and connector actions turn assistant output into usable workflow surfaces. The map makes those surfaces visible as product architecture, not just UI decoration.
What this proves
- Cross-platform assistant architecture
- Memory and connector boundaries are explicit
- Voice, cards, and automations treated as first-class surfaces
- Aura default model route plus capability-based fallbacks
17stack layers
95services mapped
40+agent tools
16app connectors
5workflow engines
3model providers
Technology stack07
Cloudflare
Chosen for a low-latency edge control plane where chat, webhooks, and connector traffic can run close to the user.
Supabase
Keeps identity, persistence, and relational memory in one Postgres-backed layer instead of scattering state across services.
Vercel AI Gateway
Carries the default Aura model route to minimax/minimax-m3 and keeps Anthropic plus Cohere rerank on the funded gateway path.
Anthropic
Keeps Opus and Sonnet available as hidden registry entries and fallback reasoning paths.
OpenAI
Covers the visible GPT-5.5 picker entry and the realtime voice model while gateway-backed chat stays the default.
Cohere
Adds retrieval and reranking strength so memory and search results can be ordered before they reach the assistant.
Apple
Fits the iOS-native side of the product: voice, app surfaces, notifications, and platform integrations.
Tools implemented06
Chat runtime
Routes user messages into the assistant, streams responses, and keeps the conversation state durable.
Interactive cards
Turns agent output into usable UI modules instead of leaving everything as plain text.
Automation engine
Lets the assistant create and manage recurring workflows, reminders, and follow-ups.
Memory layer
Stores durable facts and preferences so the assistant can act with continuity across sessions.
Connector surface
Bridges outside services into the assistant while keeping each integration isolated.
Voice interface
Adds a hands-free path for capture and response when typing is the wrong interaction.