陳鈺澔
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
I built CryptoMind, an AI agent platform for crypto and stock analysis. I have been developing this project for around 8 months.
CryptoMind can help user analyze market, ask financial questions, and also swap tokens on TON blockchain. When I started to add wallet and swap function, one thing I care a lot is I don't want LLM directly control user's money.
So for any operation related to real money, it does not run by free-form LLM action. It must go through fixed workflow. User need to confirm before execution, and the consent step cannot be skipped by agent. There is also USD amount limit, and if system cannot verify the limit correctly, it will stop the transaction instead of continue.
I also open sourced the safety rules, so people don't need to only trust what I say on the website. Server can show which safety kernel version it is using, and each swap will record a consent hash for checking later.
The platform is already running in production:
My first goal is to prove an AI agent can interact with real money in a safer way.
There are already many AI agent demos that can call tools, but when tool is related to wallet or money, I think only prompt instruction is not enough. LLM may make mistake, misunderstand user, or tool call can behave different than expected.
Because of this, CryptoMind put the money execution outside of free-form agent behavior. Swap needs to follow fixed workflow, amount limit and user confirmation. The agent cannot decide to skip these steps.
This part is already implemented and running now.
My second goal is verifiability.
CryptoMind is still an early project and users have no reason to fully trust me when they connect a wallet. So I think safety should not only be a statement like "we are secure."
I separated the important safety logic into an open-source kernel. The production server also exposes the kernel version, so the running version can be compared with source code.
For every swap, the system also generate a consent hash. It can be used as a record to show what user agreed before the transaction.
I also made a prototype to move risk checking into a Trusted Execution Environment (TEE). The prototype can generate cryptographic attestation for the execution. The prototype part is already working, but I still need to deploy it to a real TEE provider and connect it with production.
My third goal is to get an independent security audit.
Right now most of security design is still made and tested by ourselves. Even we have many automated tests, I don't think this should be enough for a system that can touch user's money.
I want an external security auditor to review the complete swap money path, including amount limit, consent flow, replay protection and transaction execution.
The platform itself is already built, so the funding is mainly for these next steps: security audit, testnet verification, supporting more token pairs, and deploying the TEE version.
The first and biggest part will be for independent security audit of the swap execution path.
I already prepared an audit scope document, including related endpoints, USD limit checking, replay protection, consent records and transaction flow. So if I can get the funding, I can directly contact auditor and start the review instead of spending funding on building the basic system first.
The second part is for testnet end-to-end verification and expanding supported token pairs.
Currently CryptoMind only support two swap pairs. I want to test more cases on testnet first, including failure case, rejected transaction and different token route, then safely add more pairs.
The third part is TEE deployment.
I already have a working prototype for TEE risk assessment and attestation. The remaining work is deploying it to a TEE provider and connecting the attestation result with CryptoMind production flow.
The goal is that execution integrity can be verified cryptographically, instead of user only trusting our backend.
I am from Taiwan and currently studying master's degree at National Taiwan University of Science and Technology (NTUST).
CryptoMind is mainly developed by myself for around 8 months. I work on most part of the system, including full-stack development, AI agent architecture, security workflow and infrastructure.
From around 1,660 commits in this project, about 1,450 are made by me.
There are also two collaborators helping me. One mainly works on market data crawling, and another helps backend and deployment.
The project started in December 2025 as a multi-agent crypto analysis system using LangGraph.
In January 2026, I changed it into a web application and added login, market data API and market screener.
From February to March, I spent more time on security and also added some community features, including forum and scam tracker. During this period I also deployed the platform to production.
In May, I finished internationalization for four languages and BYOK feature. User can bring their own LLM API key and market data API key, so the platform don't need to hold every user's API cost.
In July, I added an agent memory system and expanded the market coverage to 8 markets.
In August, I integrated TON wallet login and token swap. This is also when I added the risk-gated execution flow and open sourced the safety kernel.
Currently the full test suite has 3,161 tests and they are all passing.
I think the biggest risk is user adoption.
CryptoMind is still a very early-stage project and I don't have marketing budget. Even if technical part is working, it does not mean users will come automatically.
Another risk is external service dependency. For example, swap and TON data depend on services such as Omniston and toncenter. These services are not controlled by us, so API change, downtime or rate limit can affect CryptoMind.
If CryptoMind itself cannot get enough users, I still think some part of this project can have value.
The safety kernel is open source and MIT licensed. The design of using risk-gated workflow between AI agent and money execution can also be reused by other agent projects.
I also want to document this architecture clearly, especially why money-related tool should not only depend on LLM prompt instruction.
So even if the platform fails as a product, the safety work and implementation can still stay available for other developers.
I have not raised any external funding in the last 12 months.
The project is self-funded by me.
Until now I spent around $1,000 USD of my personal money, mainly for hosting, development tools, API testing and other infrastructure cost.
There are no bids on this project.