FinGOAT
Turning complex decisions into traceable reasoning.
A graph-orchestrated multi-agent application for US equity research. From data to view, it makes parallel reasoning, dependency control and the decision process visible.
Problem
Multi-agent systems easily fall into serial waiting: each round of analysis waits for the previous one, and users struggle to see where the final conclusion came from. FinGOAT models dependencies explicitly, runs independent analyses in parallel and keeps the full reasoning trail.
Key decisions
- Graph orchestration
- Defined the agents' execution graph and dependency control with LangGraph, so independent analyses run in parallel.
- Service boundaries
- Split responsibilities between a Gin gateway and a FastAPI inference service; PostgreSQL / GORM persists the reasoning trail and decisions, and Docker supports deployment.
- Inference flexibility
- Supports multiple APIs, local Ollama and OpenClaw for inference, leaving room to switch between models and environments.
Outcome
≈70%lower end-to-end latency than a serial pipeline
According to my project records, end-to-end decision latency in multi-agent scenarios dropped by about 70% compared with a serial pipeline. The figure is an in-project comparison; actual results depend on the model, the dependency graph and the environment.
Related notes (Chinese)