# Rick Weng > Research into how language models and AI agents actually behave, and tools that keep them accountable. - Role: Independent AI security researcher - Website: https://rickwengdev.github.io - GitHub: https://github.com/rickwengdev - LinkedIn: https://www.linkedin.com/in/rick-weng-44921a3b5 - ORCID: https://orcid.org/0009-0006-4178-347X - Eisoptric: https://eisoptric.com - Last updated: 2026-11-20 - Language: English - Also available in: https://rickwengdev.github.io/zh/ · [Markdown](https://rickwengdev.github.io/index.zh.md) ## Letter To you, I’m Rick Weng, an AI security researcher. I study how language models and AI agents actually behave, and how to keep them within the rules people set. Yours sincerely, Rick Weng ## Experience ### [Kura Finance](https://kura-finance.com) Founder & Solo Product Engineer · Jan 2026 – Sep 2026 Modern finance, redesigned. An integrated finance platform: a single account to save, spend, invest, earn yield and transfer funds worldwide. - Defined the product vision for a unified, privacy-first global finance platform. - Established partnerships with fintech infrastructure providers across digital assets, yield and payments. ## Projects ### [Eisoptric](https://eisoptric.com) Creator & Lead Maintainer · Sep 2026 – Present Infrastructure for secure, accountable AI. Eisoptric focuses on AI security: ensuring that agents operate within the rules people set, and giving teams a clear view of how models actually behave before they rely on them. - Eisoptric MCP: a gateway that enforces argument-level policy on every agent tool call, routes sensitive calls for human approval, and maintains a complete audit record. - Eisoptric App: an open-source (Apache 2.0) desktop workbench for AI research, supporting experiment management, sample-level review, activation patching and SAE feature analysis. - Credentials are held in a post-quantum vault, and audit payloads are encrypted on archival, so they remain unreadable even to the gateway itself. - Both products are free, with full functionality on every account, and are funded through sponsorship. ## Research ### Research goals - Pretraining - Interpretability --- # Rick Weng > Rick Weng 是 AI 安全研究者,研究目標為預訓練與可解釋性,探究語言模型與 AI 代理的實際行為。 - 網站:https://rickwengdev.github.io/zh/ - GitHub: https://github.com/rickwengdev - LinkedIn: https://www.linkedin.com/in/rick-weng-44921a3b5 - ORCID: https://orcid.org/0009-0006-4178-347X - Eisoptric: https://eisoptric.com - 最後更新:2026-11-20 - 語言:繁體中文 - 其他語言版本:https://rickwengdev.github.io · [Markdown](https://rickwengdev.github.io/index.md) ## 信件 To you, 我是 Rick Weng,從事 AI 安全研究,探究語言模型與 AI 代理的實際行為,以及如何讓它們在既定規則內運作。 Rick Weng 敬上 ## 經歷 ### [Kura Finance](https://kura-finance.com) 創辦人暨獨立產品工程師 · 2026 年 1 月 – 2026 年 9 月 重新設計現代金融服務。 一站式金融平台:以單一帳戶完成儲蓄、消費、投資、收益與跨境轉帳。 - 主導產品願景,打造以隱私為優先、整合全球服務的金融平台。 - 與金融科技基礎設施夥伴建立合作,涵蓋數位資產、收益與支付服務。 ## 專案 ### [Eisoptric](https://eisoptric.com) 創建者暨主要維護者 · 2026 年 9 月 – 至今 為 AI 打造安全、可問責的基礎工具。 Eisoptric 專注於 AI 安全,致力於確保 AI 代理的行為符合既定規則,並讓使用者在部署模型之前,充分理解模型的實際行為。 - Eisoptric MCP:AI 代理的工具呼叫閘道,對每一次呼叫執行參數層級的政策檢查,將敏感操作交由人工核准,並完整保存稽核紀錄。 - Eisoptric App:以 Apache 2.0 授權開源的 AI 研究桌面工作台,支援實驗管理、逐筆樣本檢視、激活修補與 SAE 特徵分析。 - 憑證儲存於後量子加密保險庫;稽核資料於歸檔時即加密,即使閘道本身也無法讀取。 - 兩項產品皆免費提供,所有帳號享有完整功能,營運經費來自贊助。 ## 研究 ### 研究目標 - 預訓練 - 可解釋性