NAV AiDE, an offline-first London travel assistantNAV AiDE，离线优先的伦敦出行助手

Your AI London guide. Works underground.你的 AI 伦敦向导，在地铁里也能用。

Ask in your own language, keep London place names grounded in English, and stay oriented when the signal drops. Built for tourists who need local guidance instead of map-only guesswork.用你自己的语言提问，伦敦地名始终锚定在英文，信号中断时也不会迷路。为需要在地指引、而不是只对着地图猜路的游客而做。

140+ languages140+ 种语言
Offline maps离线地图
Door-to-door directions门到门导航
LOST? helper迷路求助
On-device AI端侧 AI

Try the GO screen试试 GO 界面

Ask in plain language, “Waterloo to Baker Street”, or tap a suggestion. Everything routes on your device; tap the status indicator to toggle airplane mode and it keeps working.用日常语言提问，“Waterloo to Baker Street”，或点一个推荐。所有规划都在你的设备上完成；点状态指示器切到飞行模式，它照样能用。

9:41

NAV AiDE

GO, ask for a route or a placeGO，问路线或地点

Underground mode. Signal dropped, guidance stayed local.地下模式。信号中断，导航仍在本地运行。

Try试试

🧭GO

🆘LOST?

🗺️Maps地图

⚙️Settings设置

How the real app works真实应用是怎么运作的

This demo reproduces the GO screen with a simplified offline network. The shipping app goes much further.这个演示用一个简化的离线网络复刻了 GO 界面。真正发布的应用走得更远。

435-station routing435 站路网规划

Dijkstra shortest-path over a TfL-sourced graph (Tube, DLR, Elizabeth line, Overground): 575 segments, fully offline.在 TfL 官方数据构建的图上做 Dijkstra 最短路径（地铁、DLR、伊丽莎白线、Overground）：575 段线路，全程离线。

On-device AI端侧 AI

Gemma 3 (1B, GGUF via llama.rn) extracts intent to structured JSON. It never resolves locations itself.Gemma 3（1B，通过 llama.rn 跑 GGUF）把意图抽取成结构化 JSON，它从不自己解析地点。

Grounded, not guessing校准，而非猜测

An EntityResolver matches every place against local indices first; low confidence shows disambiguation, not a hallucination.EntityResolver 先把每个地点对照本地索引匹配；置信度低时弹出消歧选项，而不是凭空编造。

860 POIs and voice860 个地点与语音

Museums, galleries and landmarks via SQLite FTS5; voice input in 140+ languages with OS-native speech and TTS.博物馆、画廊、地标通过 SQLite FTS5 检索；140+ 种语言语音输入，调用系统原生语音识别与 TTS。

React Native + TypeScriptOffline-first离线优先On-device LLM端侧 LLM

Concept and product by me; scaffolding and iteration done by directing AI coding agents. The demo above is a faithful re-creation of the GO screen's design and flow, with a small illustrative network so it runs in any browser.概念与产品由我主导；代码脚手架与迭代通过指挥 AI 编程智能体完成。上方的演示忠实复刻了 GO 界面的设计与流程，用一个小型示意网络，让它能在任意浏览器里运行。

See the open-source repo on GitHub在 GitHub 上查看开源仓库

NAV AiDE, concept and open-source productNAV AiDE，概念与开源产品 Back to the work返回作品