Rezon

Inspiration

Most trading tools focus on charts, numbers, and automation, but few address the trader behind the screen. Every trader has habits: selling too early, holding risky stocks too long, or chasing hype. We wanted to build a system that helps traders understand and correct those habits in real time through an intelligent overlay that reacts to their actions and trading style.

What it does

Rezon is a desktop overlay built with Electron that acts as a personal trading companion. It connects to your browser through Chrome Remote Debugging and interprets what you're doing, such as hovering over a "Buy" button, viewing a chart, or researching new stocks. Based on your personalized trading profile, it provides timely insights, warnings, and recommendations that align with your individual trading approach.

You start by defining your trading personality: what kind of trader you are and what you struggle with. From there, Rezon observes your activity and builds short, meaningful contexts for a local AI agent powered by the Qwen2 model. This agent uses an internal graph of "agentic nodes" to reason through your behavior and determine which tools to activate, such as YFinance for live financial data or DuckDuckGo for relevant stock and company news. It operates entirely on-device, allowing for fast, privacy-preserving decisions that evolve dynamically with your actions. At the end of your session, Rezon generates a concise summary of your trades, insights, and behavioral patterns, helping you reflect and improve over time.

How we built it

The frontend was built using Electron, which powers the desktop overlay and user interface. For browser integration, we used the Chrome Remote Debugging Protocol with WebSocket communication to monitor DOM changes and user actions in real time. We designed an event system using a lightweight poller that detects meaningful user interactions and converts them into structured event contexts for the AI to interpret.

The backend was built entirely in Python. It manages the agentic node system and runs the Qwen2 large language model locally to ensure fast, secure, and explainable reasoning. Each node represents a step in the AI's reasoning process, deciding whether to pull market data from YFinance, perform a web search through DuckDuckGo, or generate feedback for the user. This architecture allows the agent to self-direct its workflow without relying on cloud inference, combining autonomy with accountability.

Challenges we ran into

We had to engineer a stable real-time connection between the Electron overlay and Chrome's debugging interface without interrupting normal browsing performance. Another challenge was optimizing the communication loop between the frontend, the local Qwen2 model, and the Python node graph to maintain smooth reasoning and rapid feedback without lag. We also worked on designing a meaningful and efficient schema for passing DOM updates and user context to the AI agent, ensuring that it receives actionable and concise information without unnecessary noise.

Accomplishments that we're proud of

We created a working prototype of a locally hosted, agentic trading companion that meaningfully augments human decision-making rather than automating it. Rezon demonstrates how an on-device LLM can actively interpret, reason, and assist in real time without sacrificing user privacy. The agent's dynamic node graph and contextual awareness allow it to provide relevant and trustworthy guidance as the user interacts with the market.

What's next for Rezon

We plan to refine the Qwen2 reasoning layer, improve real-time responsiveness, and expand Rezon to support multiple trading platforms and brokers. We also aim to integrate deeper analytical nodes capable of simulating risk scenarios or modeling emotional patterns in trades. Our long-term goal is to build a personalized trading assistant that helps users understand why they make certain decisions and how to make better ones.