🚴‍♂️ AI Cycling Companion — Liquid AI Hackathon Project (2nd Place 🥈)

🏆 2nd Place — Liquid AI x W&B x Lambda Hackathon Tokyo

---

🌟 Overview

This project was built for the Liquid AI x W&B x Lambda Hackathon in Tokyo 🇯🇵,
where our team proudly won 2nd place 🥈 for our creative concept.

Idea:
A local AI cycling companion that speaks Japanese naturally while you ride —
like a friendly character from Cars or Thomas the Tank Engine.
It’s designed to make cycling more emotional, conversational, and fun 🚲💬

---

💡 Concept Flow

• 🎙️ Record short voice clips via the mobile app
• 🧩 Transcribe audio (Whisper / ASR)
• 🐍 Send text to a FastAPI backend connected to LFM2-VL (Japanese multimodal LLM)
• 🤖 Model replies in natural Japanese
• 🔊 Response is spoken aloud — creating a real-time “AI riding buddy”

---

🧰 Tech Stack

Layer	Description
Frontend	Expo (React Native) — voice capture & UI
Backend	FastAPI — transcription + model bridge
Model	LFM2-VL — multimodal Japanese LLM
Infra	Lambda Cloud — training and fine-tuning
Tracking	Weights & Biases — experiments and evaluation
Language Focus	🇯🇵 Japanese — for natural, culture-aware interactions

---

🧩 Architecture Overview

sequenceDiagram;
    participant U as 🧍 User (Cyclist);
    participant A as 📱 Expo App (React Native);
    participant S as 🐍 FastAPI Server;
    participant M as 🤖 LFM2-VL Model;

    U->>A: Press "Talk" button;
    A->>A: Record short audio;
    A->>S: Send audio file + prompt;
    S->>S: Transcribe via Whisper;
    S->>M: Send text to LFM2-VL;
    M->>S: Return Japanese response;
    S->>A: Return JSON reply;
    A->>U: Display + speak reply aloud;

Loading

---

🏆 Recognition

This project stood out for its creativity and human-centered design — bringing together Japanese culture, empathy, and conversational AI.

“It feels like riding with a character who actually understands you.”

---

🙏 Acknowledgements

Huge thanks to everyone who made this project possible:

• Liquid AI Tokyo — Hackathon organizers
• Lambda Cloud — Model training environment
• Weights & Biases (W&B) — Experiment tracking tools
• @Hayato Hongo — LFM2-VL model author & fine-tuning
• @Rikka Botan — Model deployment and integration

---

📘 Related Links

• Event Guide (Notion)
• Liquid AI Hugging Face Models
• Hayato Hongo LFM2-VL Repo
• Rikka Botan GGUF Model

---

🚀 Future Work

• Real-time conversation & latency optimization
• Add emotional tone and “ride personality”
• Expand to regional Japanese dialects (関西弁, 東北弁, etc.)

---

🇯🇵 Built in Tokyo — where empathy meets innovation.

“Vibe coding meets human connection.”

---

概要（日本語）

このプロジェクトは、Liquid AI × W&B × Lambda Hackathon Tokyo 🥈 で 2 位を受賞した「AI 自転車コンパニオン」アプリです。
React Native（Expo）で開発されたモバイルアプリと、FastAPI サーバー上で動作する日本語マルチモーダルモデル（LFM2-VL）を組み合わせ、
ユーザーが自転車に乗りながら自然な日本語で会話できる体験を実現しました。

日本の「共感」と「創造性」を活かした、人間中心の AI インタラクションを目指しています。
開発・運用には Hugging Face, Lambda Cloud, Weights & Biases を使用しました。