Lantern Edge — Architecture

Offline-first edge AI agent · Real Qwen models on-device (WebGPU), graceful degradation, cross-device sync

🖥️ Browser (Client / Edge Device)

🧠

WebLLM Engine

Qwen2.5 0.5B–3B
runs via WebGPU
0 server calls

🧮

On-Device Tools

Time · Calculator · Notes
No cloud needed
Works fully offline

⏳

Deferred Queue

Stores tasks when offline
Auto-executes on reconnect
Survives disconnects

↑↓ user input · streaming tokens · tool results

🧭 Routing Layer (Browser — Model-Free Classifier + Degradation Policy)

📊

Task Classifier

Scores complexity, capabilities,
privacy & reasoning depth
(no LLM call needed)

🔀

Degradation Policy

decideRoute(): edge-first →
escalate to cloud → or
defer to queue

→ Route A (edge) · Route B (cloud) · Route C (defer)

⚙️ Express Server (Node.js — secrets stay server-side)

☁️

Qwen Cloud Proxy

DashScope streaming (SSE)
qwen-turbo / plus / max
API key never leaves server

🔄

OSS Sync

Alibaba Cloud OSS
Conversation + state sync
Local-disk fallback

→ external API calls

☁️ External Cloud Services

🤖

Qwen Cloud (DashScope)

dashscope-intl.aliyuncs.com
OpenAI-compatible API
Streaming completions

📦

Alibaba Cloud OSS

Bucket: lantern-edge-sync
Cross-device memory
Conversation persistence

🔀 The Four Routing Paths

Edge (Offline)

Simple tasks → handled by on-device WebGPU model + local tools. Zero network calls. Works fully offline.

Cloud (Online)

Complex tasks → escalate to Qwen Cloud via server proxy. API key stays server-side. Streams tokens back to browser.

Deferred (Offline + Complex)

Hard task while offline → stored in deferred queue. Auto-executes via Qwen Cloud when connection restores.

Sync (On Reconnect)

Conversations + agent state synced to Alibaba Cloud OSS. Pick up on any device. Local-disk fallback when OSS unavailable.

📊 Project Metrics

35

Tests passing

5

Qwen edge models

4

Routing paths

100%

Offline-capable core

🔥 Lantern Edge — Architecture