Architecture Overview

Anky runs as a single Rust binary on a bare-metal machine called poiesis. No microservices. No containers in production. No cloud databases.

System Diagram

┌──────────────────────────────────────────────────────────────┐
│  poiesis (bare metal)                                        │
│                                                              │
│  ┌─────────────┐  ┌──────────┐  ┌────────────────────────┐  │
│  │ Anky (Rust)  │  │  Ollama  │  │  ComfyUI + Flux LoRA  │  │
│  │  port 8889   │  │  :11434  │  │       :8188            │  │
│  │              │  │          │  │                        │  │
│  │ Axum router  │──│ qwen3.5  │  │  flux1-dev +           │  │
│  │ SQLite DB    │  │  :35b    │  │  anky_flux_lora_v2     │  │
│  │ Tera tmpl    │  │          │  │                        │  │
│  └──────┬───────┘  └──────────┘  └────────────────────────┘  │
│         │                                                     │
│         │ systemd: anky.service                               │
└─────────┼────────────────────────────────────────────────────┘
          │
          │ cloudflared tunnel
          │
    ┌─────▼─────┐
    │ Cloudflare │
    │  anky.app  │
    └───────────┘

Components

Rust/Axum Server

• Single binary: target/release/anky
• HTTP router with Axum
• Server-rendered HTML via Tera templates
• Mobile API under /swift/v1/* and /swift/v2/*
• Background workers spawned as tokio tasks

SQLite

• Single file: data/anky.db
• Accessed via rusqlite through AppState
• Schema managed by src/db/migrations.rs (not external SQL files)
• No ORM — raw SQL queries in src/db/queries.rs

Ollama (Local Text AI)

• Model: qwen3.5:35b
• URL: http://localhost:11434
• Used for: free-tier guidance, image prompts, writing feedback, translations
• Data never leaves the machine

ComfyUI (Local Image AI)

• Flux.1-dev with Anky LoRA (strength 0.85)
• URL: http://127.0.0.1:8188
• Square images: 1024x1024 (ankys)
• Vertical images: 768x1344 (cuentacuentos phases)
• Sampling: 20 Euler steps, CFG 3.5

Claude (Cloud Text AI)

• Used for: premium guidance, chakra detection, cuentacuentos generation, facilitator matching
• Accessed via API with payment verification
• Never touches the write path directly

Cloudflare Tunnel

• Service: cloudflared-anky.service
• Routes anky.app → localhost:8889
• No exposed ports, no load balancer, no reverse proxy config

Data Flow

User writes 8+ min
       │
       ▼
POST /swift/v2/write
       │
       ├── Persist to SQLite (writing_sessions)
       ├── Mirror to data/writings/{wallet}/{ts}.txt
       │
       ├── Spawn: Anky Image Pipeline
       │   └── Ollama prompt → Gemini/Flux → WebP + thumb
       │
       │
       └── Spawn: Cuentacuentos Pipeline (v2 only)
           ├── Claude → chakra + story + phases
           ├── Ollama → image prompts per phase
           ├── Flux → vertical images per phase
           └── Ollama → translations (es, zh, hi, ar)

Why This Architecture?

• Single binary: No deployment coordination. cargo build --release && systemctl restart.
• SQLite: No database server. No connection pooling. No migrations infra. Just a file.
• Bare metal: GPU access for Flux. No container overhead. No cloud GPU costs.
• Local AI: Free-tier costs $0. No API rate limits. No data exfiltration.
• Cloudflare tunnel: Zero-config HTTPS. No certificate management. No port forwarding.