AI Decision Engine: How Multi-Agent Architecture Turns Major Life Decisions into Structured Analysis

This article introduces an open-source multi-agent AI decision engine that transforms complex life/career decisions into 11-dimensional structured analysis through risk modeling, regret minimization, and strategic frameworks, helping decision-makers in high-pressure environments make rational choices. The core of the project is a three-agent collaborative architecture, combined with visualization and quantitative tools, to provide decision support beyond intuition.

Author Dharshan Sondi, dissatisfied with the vague advice (e.g., "follow your passion") given by traditional AI for major decisions (such as quitting a job to start a business or choosing a job), hoped that AI could analyze deeply like professional roles such as venture capitalists and strategists. This system is not a simple LLM wrapper but a structured cognitive architecture that enforces structured reasoning by revealing hidden assumptions, mapping opportunity costs, quantifying risks, etc.

The core of the system is three-agent collaboration:

• Visionary: Maximize upside, explore leverage effects, options, and asymmetries;
• Risk Manager: Protect downside, detect existential risks, vulnerabilities, and hidden traps;
• Synthesizer: Integrate outputs from the previous two to generate an 11-dimensional analysis report. This design simulates a decision committee to avoid blind spots from a single perspective.

Each decision is evaluated through 11 dimensions: Problem Framing (clarify core and assumptions), Constraint Mapping (financial/geographical restrictions, etc.), Risk Analysis (quantify severity), Opportunity Cost (hidden costs), Skill Gap (comparison between existing and required skills), Strategic Paths (probability and reversibility of multiple paths), Probability Model (short/medium/long-term outcome distribution), Recommendation Set (multi-perspective suggestions), Cognitive Bias Detection (identify thinking traps), Antifragility Score (adapt to uncertainty), Regret Minimization (analysis from an 80-year-old's perspective).

The system solves several technical problems:

1. JSON Repair Mechanism: Reverse traversal to fix truncated JSON, ensuring valid data is received by the frontend;
2. PDF Generation Optimization: Isolate logic into a dedicated module, use sync_playwright and thread isolation to resolve Windows environment errors;
3. Rate Limiting and Quota Management: IP-based token bucket limiting, support for Bring Your Own Key (BYOK), track AI health status and prompt when quota is exhausted.

The system provides rich interactions:

• Data Visualization: Risk radar chart, skill gap radar chart, path comparison bar chart, React Flow decision tree (auto-layout, zoom);
• Real-time Features: Animated gauges showing confidence/antifragility, etc.;
• Export and Sharing: Client/server-side PDF export, Markdown export, shareable links;
• Others: Text-to-speech (multilingual), analysis history, comparison mode.

Applicable to founders, engineers, and others facing high-risk choices, such as quitting to start a business, choosing a job/degree, career transition, etc. Unlike traditional vague advice, the system provides structured analysis (e.g., "Path A has a 45% upside probability but an 8/10 existential risk"), acting as a decision architecture rather than a conversation, and computation rather than advice.

AI Decision Engine demonstrates the application of multi-agent architecture in personal decision support, helping to go beyond emotional decisions. Future plans include: support for multiple AI providers (OpenAI, Anthropic), user authentication and cloud history, team collaboration analysis, mobile PWA, real-time streaming responses, decision result tracking and feedback. The project is open-source; developers are welcome to contribute PRs or Issues.