Zelara Core

Zelara Core — AI Hub

Status: v0.6.0 Phase 5 complete (2026-04-08) — AI runtime platform, PaddleOCR ONNX pipeline, Qwen3 persistent subprocess, model manager, hardware tier detection, finance_import.rs

Zelara Core is the desktop application that powers every Zelara app. It runs on your computer and exposes local AI capabilities over an encrypted connection to your phone. Every Zelara mobile app can connect to it and offload heavy computation — receipt scanning, language model inference, image recognition, document parsing — without sending data to any cloud.

Tagline: "Your personal AI hub. Install once, power all your Zelara apps."

Why It Exists

Most "AI-powered" apps run models in a data center. That means:

Your photos and documents leave your device
You pay a subscription for someone else's compute
The app stops working when the server is down or the company changes pricing

Zelara Core inverts this. Your desktop has significantly more compute than your phone. Models that would be impractical on a phone (PaddleOCR, DistilBERT, local LLMs) run comfortably on a modern laptop. Zelara Core manages those models and serves results back to your phone over a TLS-encrypted local connection — same network, milliseconds of latency, zero cloud.

What It Does

1. AI Model Manager

Zelara Core maintains a catalog of ONNX and GGUF models used by Zelara apps:

Model	Purpose	Size	Used By
PaddleOCR v5 (det + rec)	Receipt scanning, document OCR	~20 MB	Zelara Finance
distilbert-base-multilingual-cased	Transaction auto-categorization	~65 MB	Zelara Finance
MobileNet v3 (ONNX)	Recycling CV — bag/recyclables detection	~10 MB	Zelara (main)
Mistral 7B GGUF (optional)	Natural language spend coaching, Q&A	~4 GB	Zelara Finance (enhanced)

Core downloads models on demand (first time an app requests a capability that needs them), caches them to the app data directory, and reuses them across sessions. Models are never re-downloaded unless explicitly updated by the user.

2. App Connection Hub

Any Zelara app on the local network can connect to Core via:

BLE auto-discovery: Core advertises its IP:port over Bluetooth at launch; apps connect automatically when in range — no user action required
QR pairing: Core displays a QR code for manual first-time pairing on unfamiliar networks

Once connected, the app gets:

AI task results (OCR, categorization, CV, LLM inference)
Progress sync — Core is the authoritative source of truth for skill tree state
Cross-app data aggregation (e.g., Finance totals surfaced in Core's dashboard)

3. AI Task Dispatcher

Core receives ai_task messages over WSS and routes them to the right model:

{
  "type": "ai_task",
  "task_id": "uuid",
  "capability": "ocr_receipt | categorize_transaction | classify_image | llm_query",
  "payload": {}
}

Response:

{
  "type": "ai_task_result",
  "task_id": "uuid",
  "result": {},
  "processing_ms": 340
}

This replaces the current hardcoded image_validation task type with a general-purpose capability dispatch system. Each capability maps to a Rust handler in src-tauri/src/ai/.

If the required model is not yet downloaded, Core responds with:

{
  "type": "ai_task_result",
  "task_id": "uuid",
  "status": "model_required",
  "model": "paddleocr-v5",
  "size_mb": 20
}

The app then surfaces a "Download model on Core" prompt to the user.

4. Finance Data Mirror

When Zelara Finance is connected, Core receives a finance_expense_sync push whenever a new expense is logged. Core persists this in its own SQLite Finance DB and surfaces it in the full-screen desktop Finance dashboard. This follows the same broadcast pattern as progress_sync (already implemented in device_linking.rs).

5. Progress Authority

Core (storage.rs) is the authoritative source for user skill tree state:

Points earned across all connected Zelara apps accumulate here
Unlock thresholds checked and enforced server-side
progress_sync broadcast to all connected apps after any state change

UI/UX Redesign

Current state (Zelara Desktop, pre-v0.6.0)

Header: "Zelara Desktop"
Sections: Device Linking / Testing Panel / Module List
Developer-tool aesthetic — QR code as hero UI, status badges, debug log in center view
Not useful to non-developer users

Target state (Zelara Core)

Four-section side navigation. Designed for visibility and control — a power user who wants to know what their AI hub is doing, not just a developer tool.

Hub (landing / home view)

┌──────────────────────────────────────────────────────┐
│  Zelara Core        ● AI Hub active   [Settings ⚙]   │
├──────────────────────────────────────────────────────┤
│                                                      │
│  Connected Apps                                      │
│  ┌────────────────────┐  ┌────────────────────┐     │
│  │ 💰 Zelara Finance   │  │ 📱 Zelara           │     │
│  │ ● Connected         │  │ ○ Not linked        │     │
│  │ 14 tasks today      │  │                     │     │
│  └────────────────────┘  └────────────────────┘     │
│                                                      │
│  Today's Activity                                    │
│  ┌──────────────────────────────────────────────┐   │
│  │ 09:42  Receipt OCR (340ms)   Zelara Finance  │   │
│  │ 09:41  Categorization ×3     Zelara Finance  │   │
│  │ 09:38  Recycling validation  Zelara (main)   │   │
│  └──────────────────────────────────────────────┘   │
│                                                      │
│  Network                                             │
│  ● WSS server running on :8765                       │
│  ● BLE advertising 192.168.1.42:8765                │
│  1 device connected                                  │
│                                                      │
│  [Pair New Device]                                   │
└──────────────────────────────────────────────────────┘

Key changes from current UI:

"Zelara Desktop" → "Zelara Core"
App cards replace abstract "device count" — user sees which named apps are connected
Activity feed shows real AI tasks processed with timing (not just a developer debug log)
QR pairing is a secondary action ([Pair New Device] button), not the hero UI

AI Models

┌──────────────────────────────────────────────────────┐
│  AI Models          Storage: 95 MB used              │
├──────────────────────────────────────────────────────┤
│                                                      │
│  Installed                                           │
│  ─────────────────────────────────────────────────  │
│  ● PaddleOCR v5              20 MB   9× today        │
│    Receipt & document OCR (80+ languages)            │
│                                                      │
│  ● DistilBERT (multilingual)  65 MB  3× today        │
│    Transaction auto-categorization                   │
│                                                      │
│  ● MobileNet v3 (CV)          10 MB  1× today        │
│    Recycling bag detection                           │
│                                                      │
│  Available to Install                                │
│  ─────────────────────────────────────────────────  │
│  ○ Mistral 7B GGUF            4.1 GB  [Download]     │
│    Enhanced spend coaching and natural language Q&A  │
│    Requires: 8 GB+ RAM, ~15 min first download      │
│                                                      │
└──────────────────────────────────────────────────────┘

Models are stored at:

Windows: %APPDATA%\Zelara\models\
macOS/Linux: ~/.zelara/models/

Download progress is shown inline. Models load into memory on first use and stay resident while Core is running.

Finance Dashboard (shown when Zelara Finance app is connected)

┌──────────────────────────────────────────────────────┐
│  Finance      [Import ▼] [Export ▼]     April 2026   │
├──────────────────────────────────────────────────────┤
│                                                      │
│  ┌──────────────┐  ┌──────────────┐  ┌────────────┐ │
│  │ Total Spent  │  │ Transactions │  │  Surplus   │ │
│  │  $1,247.83   │  │     47       │  │  $752.17   │ │
│  └──────────────┘  └──────────────┘  └────────────┘ │
│                                                      │
│  Spending Trend (6 months, Recharts animated line)   │
│  Nov — Dec — Jan — Feb — Mar — Apr(forecast····)     │
│                                                      │
│  By Category              Top Merchants              │
│  Food       ████ $423     Walmart    $287             │
│  Transport  ███  $187     Shell      $124             │
│  Utilities  ███  $164     Starbucks   $89             │
│                                                      │
│  Recent Transactions                                 │
│  Apr 4  Walmart  Food    -$45.23  Receipt OCR        │
│  Apr 4  Shell    Trans   -$62.00  Manual             │
│  [See all →]                                         │
│                                                      │
│  [📁 Import Bank File]     [📤 Export Report]         │
└──────────────────────────────────────────────────────┘

When no Finance app has ever connected: show "Install Zelara Finance on your phone to see your financial dashboard here."

Drag-and-drop file import: drop OFX / CSV / XLSX / PDF onto the window → auto-detected format → Rust backend parses via finance_import.rs → preview parsed rows → confirm.

Settings

  Network
  ───────────────────────────────────────────────
  WSS Port                    8765
  BLE Advertising             ● On    [Toggle]
  TLS Certificate             Self-signed (generated 2026-02-15)
                              [Regenerate Certificate]

  Storage
  ───────────────────────────────────────────────
  Models directory            %APPDATA%\Zelara\models\    [Change]
  Finance database            %APPDATA%\Zelara\finance.db
  Progress file               %APPDATA%\Zelara\progress.json

  Connected Apps (pairing history)
  ───────────────────────────────────────────────
  Zelara Finance              Last seen: today 09:42
  Zelara (main)               Last seen: today 09:38
  [Forget all paired devices]

  Debug
  ───────────────────────────────────────────────
  Points: 60    [+10] [+50] [+9999] [Reset to 0]

Debug controls moved here from the old TestingPanel — keeps the hub views clean.

Logic Redesign (Rust Backend)

New file structure in `apps/desktop/src-tauri/src/`

src-tauri/src/
├── lib.rs                    (command registration — updated)
├── storage.rs                (progress/unlock — unchanged)
├── device_linking.rs         (WSS server, broadcast — extended for ai_task)
├── ble_advertising.rs        (unchanged)
├── cv_processor.rs           (recycling CV — now delegates through ai/ dispatcher)
├── finance_import.rs         (desktop-side file parsing: calamine, csv-core, ofxy)
└── ai/
    ├── mod.rs                (capability registry + re-exports)
    ├── dispatcher.rs         (routes AiTask to the right handler by capability string)
    ├── model_manager.rs      (download, cache, load ONNX/GGUF models to app_data_dir)
    ├── ocr_receipt.rs        (PaddleOCR v5 — det + rec pipeline, vertical tiling for long receipts)
    ├── categorization.rs     (DistilBERT tokenize + ONNX inference → Category)
    └── llm.rs                (llama-cpp-rs bindings for Mistral 7B — feature-flagged: "llm")

WSS message types added to `device_linking.rs`

"ai_task" => {
    let task: AiTask = serde_json::from_value(payload)?;
    let result = ai::dispatcher::dispatch(task, &app_handle).await;
    send_to_client(client_id, result).await;
}

"finance_expense_sync" => {
    let expense: ExpenseRecord = serde_json::from_value(payload)?;
    finance_import::record_expense(&expense, &app_handle).await?;
    app_handle.emit("finance-expense-received", &expense)?;
}

`AiTask` struct

#[derive(Deserialize)]
pub struct AiTask {
    pub task_id: String,
    pub capability: String,   // "ocr_receipt" | "categorize_transaction" | "classify_image" | "llm_query"
    pub payload: serde_json::Value,
}

#[derive(Serialize)]
pub struct AiTaskResult {
    pub task_id: String,
    pub result: Option<serde_json::Value>,
    pub status: String,       // "ok" | "model_required" | "error"
    pub model: Option<String>,
    pub size_mb: Option<u32>,
    pub processing_ms: u64,
}

Model Manager behavior

Models stored at {app_data_dir}/models/{model_name}/
get_model_status Tauri command → returns list of installed models + storage bytes (for Models UI)
download_model(name) Tauri command → streams download, emits model-download-progress events to frontend
On ai_task with missing model → immediate model_required response (no hanging)
Loaded models cached in Arc<Mutex<HashMap<String, OnnxSession>>> in Tauri state

`cv_processor.rs` integration

The existing image_validation task type in device_linking.rs stays as-is for backward compatibility with the current Zelara mobile app. Internally, it will route through ai::dispatcher::dispatch once the ai/ module exists, but the external WSS message type remains image_validation so no mobile app update is required to maintain recycling validation.

Migration from Current Desktop UI

Current component	What happens in redesign
`DevicePairing.tsx`	Absorbed into Hub view — "Pair New Device" button triggers QR modal
`ProgressDisplay.tsx`	Removed from header; compact points badge only (e.g., "60 pts")
`ModuleList.tsx`	Removed — "modules" concept replaced by connected app cards
`TestingPanel.tsx`	Retained for dev use; debug point controls move to Settings
New: `HubView.tsx`	Main landing: app cards, activity feed, network status
New: `ModelsView.tsx`	AI model catalog, download, storage usage
New: `FinanceDashboard.tsx`	Full-screen Finance analytics (Finance app must be connected)
New: `SettingsView.tsx`	Network config, BLE toggle, storage paths, debug controls

App.tsx router:

type View = 'hub' | 'models' | 'finance' | 'settings';

Side nav layout (icon + label, minimal):

💻  Hub
🧠  AI Models
💰  Finance
⚙️  Settings

Capability Roadmap

Capability	v0.6.0	v0.7.0	v1.x
Recycling CV (MobileNet)	✓ existing	✓	✓
Receipt OCR (PaddleOCR v5)	✓	✓	✓
Transaction categorization (DistilBERT)	✓	✓	✓
Rule-based spend coaching	✓	✓	✓
Finance desktop dashboard	✓	✓	✓
Local LLM spend coaching (Mistral 7B)	deferred	✓	✓
Document intelligence (general)	—	✓	✓
Health/fitness CV models	—	—	✓

Relationship to Other Zelara Apps

Zelara Core (Desktop)
├── powers → Zelara Finance (receipt OCR, categorization, finance dashboard)
├── powers → Zelara (main app) (recycling CV, progress sync)
└── powers → [future apps] (any AI capability they need)

Zelara Finance → earns points → syncs to Core → Core broadcasts → all connected apps
Zelara (main) → earns points → syncs to Core → Core broadcasts → all connected apps

Core is never a mobile app. It is always the desktop hub. Mobile apps work in degraded mode (rule-based fallbacks) when Core is not connected, and get full AI capabilities when connected.

Navigation Menu

Uh oh!

Zelara Core

Zelara Core — AI Hub

Why It Exists

What It Does

1. AI Model Manager

2. App Connection Hub

3. AI Task Dispatcher

4. Finance Data Mirror

5. Progress Authority

UI/UX Redesign

Current state (Zelara Desktop, pre-v0.6.0)

Target state (Zelara Core)

Hub (landing / home view)

AI Models

Finance Dashboard (shown when Zelara Finance app is connected)

Settings

Logic Redesign (Rust Backend)

New file structure in `apps/desktop/src-tauri/src/`

WSS message types added to `device_linking.rs`

`AiTask` struct

Model Manager behavior

`cv_processor.rs` integration

Migration from Current Desktop UI

Capability Roadmap

Relationship to Other Zelara Apps

See Also

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Clone this wiki locally

Uh oh!

Zelara Core

Zelara Core — AI Hub

Why It Exists

What It Does

1. AI Model Manager

2. App Connection Hub

3. AI Task Dispatcher

4. Finance Data Mirror

5. Progress Authority

UI/UX Redesign

Current state (Zelara Desktop, pre-v0.6.0)

Target state (Zelara Core)

Hub (landing / home view)

AI Models

Finance Dashboard (shown when Zelara Finance app is connected)

Settings

Logic Redesign (Rust Backend)

New file structure in apps/desktop/src-tauri/src/

WSS message types added to device_linking.rs

AiTask struct

Model Manager behavior

cv_processor.rs integration

Migration from Current Desktop UI

Capability Roadmap

Relationship to Other Zelara Apps

See Also

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Clone this wiki locally

New file structure in `apps/desktop/src-tauri/src/`

WSS message types added to `device_linking.rs`

`AiTask` struct

`cv_processor.rs` integration