Growing Local Voices in Native LLMs: Building the "Starlight Zerobot" with Twinkle AI 4B
- Time
- 2026-08-09 10:40 ~ 11:10
- Speaker
- Haley
- Room
- TR212
- Co-write
Abstract
ZeroTree is a gamified goal-management project combining psychology and skill trees, initially created to overcome my own procrastination. This session follows my journey from a 2-week extreme integration challenge — bringing Twinkle AI's Taiwan-native 4B model (gemma-3-4B-T1-it) into ZeroTree from scratch — through months of continuous iteration. Centered around a "mood diary," I developed the "Starlight Zerobot," an AI companion prototype featuring Taiwanese open-source community humor and local context. I will dive deep into four major challenges and solutions of deploying a local Small Language Model (SLM): JSON formatting constraints, System Prompt optimization for CPU inference caching, UX design to mitigate hardware latency, and anonymous privacy protection for local projects.
Speaker
Haley
我是一名 AI 工程師，目前專注於 LSTM 時序預測的開發。我具備電機與腦科學背景，研究所期間專攻利用 fMRI 記錄腦部狀態，並嘗試透過 GAN 技術重建受試者當下的視覺情景。 由於自身深為拖延症與思緒過載所苦，我發起了「零樹計畫（ZeroTree）」，期望藉由遊戲化技能樹的設計來輔助自我成長。專案初期，我將《實現：達成目標的心智科學》 一書中的心理學概念實作為目標設立系統；近期更挑戰在兩週的極短時間內從零開始，將 Twinkle AI 台灣原生模型落地整合進專案，打造出以「心情日記」為核心的在地化 AI 陪伴系統。我熱衷於將生硬的技術結合心理學，轉化為有溫度的日常實用工具。