AlphaCouncil: Multi-Agent Debate-Driven Taiwan Stock Analysis with Google ADK + Open-Source LLMs — Localizing Both Data and Models
- Time
- 2026-08-09 14:00 ~ 14:30
- Speaker
- Samson, 陳庭宇
- Room
- AU
- Co-write
Abstract
Open-source projects like ai-hedge-fund (⭐ 52k stars) and TradingAgents (⭐ 50k stars) have proven the community's appetite for LLM-powered financial analysis. But both projects are US-centric—their data sources, news feeds, and financial metrics are built around American markets and don't transfer cleanly to Taiwan.
This talk introduces AlphaCouncil, a Taiwan-first multi-agent stock analysis framework built natively on Google ADK, combining investor persona agents (inspired by ai-hedge-fund) with debate-driven decision-making (inspired by TradingAgents). Localization happens at two levels:
Data layer — replacing US-centric sources with Taiwan equivalents:
News: Traditional Chinese financial media (Yahoo Finance TW, 鉅亨網, 經濟日報, etc.) Institutional flows: TWSE OpenAPI for 三大法人 (foreign investors, investment trusts, dealers) and margin trading (融資融券)—free, no API key required Fundamental metrics: yield (殖利率), monthly revenue (月營收), P/B ratio—Taiwan-specific KPIs absent from US-oriented feeds Model layer — running open-source LLMs locally via Ollama (Gemma 4, Llama 3, Qwen, etc.):
All 13 investor persona reports generated in Traditional Chinese with Taiwan market context Financial data never leaves the local machine—addressing privacy requirements in financial use cases Gemma 4 pairs naturally with Google ADK (same ecosystem), delivering strong Traditional Chinese instruction-following and multi-step reasoning out of the box No commercial API costs; commercial APIs (Gemini, Claude, GPT) remain an optional swap-in The pipeline runs in five stages—analyst team, masters panel, research debate, trader, and risk debate—orchestrated with SequentialAgent, ParallelAgent, and LoopAgent primitives to ensure a deterministic flow without LLM dynamic routing.
Key engineering takeaways:
Integrating Gemma 4 via Ollama into a Google ADK multi-agent pipeline Integrating TWSE's free OpenAPI into an ADK tool layer Prompt engineering for Traditional Chinese persona-constrained output with Taiwan market context
Speaker
Samson
現任國泰金控助理架構師，擁有 9 年以上系統架構與後端開發經驗，橫跨金融、政府、醫療與車用等多個產業領域。目前職責著重於企業 AI 賦能與架構治理，推動內部 AI 導入策略與工程實踐。
長期關注開源技術在垂直領域的落地應用，熟悉雲端部署與分散式系統設計。
本次議程以自行實作的 AlphaCouncil 開源專案為主軸，分享在金融場景中導入 Google ADK 與開源 LLM（Gemma 4）構建多代理人系統的工程實踐。
Assistant Architect at Cathay Financial Holdings with 9+ years of system architecture and backend development experience across financial services, government, healthcare, and automotive domains. Currently focused on enterprise AI adoption and architecture governance.
A practitioner interested in applying open-source technology to real-world vertical domains, with background in cloud deployment and distributed systems design.
This talk is grounded in hands-on experience building AlphaCouncil: an open-source multi-agent Taiwan stock analysis framework integrating Google ADK with locally-deployed open-source LLMs (Gemma 4).