LangGraph Financial Swarm
投研报告动辄几十页、引用链路散落十几个文件，谁来读、谁来核、结论怎么追溯？五个 AI 专家分头扒数据、交叉验证、最后合成可追溯的结论：宏观分析师讲周期、行业研究员拆赛道、公司专家读财报、风险官挑漏洞、PM 把结论落到投资决策。每一段结论都标注原始出处和推理路径，监管和合规可以直接审计。整套系统用 LangGraph 编排 + 本地 LLM，企业内网就能跑，数据不出域。适合券商研究所、买方投研、家族办公室，把分析师从 60% 的重复劳动里解放出来，专注真正的判断。
Investment research needs traceable specialist reasoning, but a single model answer often mixes assumptions, sources, and risk judgments into an opaque narrative.
I modeled the workflow as a LangGraph state machine with specialist nodes, shared state, checkpointing, and a final arbiter step.
The project demonstrates how financial analysis can become replayable, branchable, and easier to audit when agent roles are explicit.
- 01Modeled the specialist-agent workflow and shared state transitions.
- 02Built the local-first Python orchestration flow.
- 03Focused the architecture on replayability, isolation, and analyst trust.
Financial research is a conversation between specialists, not a single prompt — a LangGraph state machine orchestrates five agents (macro, equity, fundamental, sentiment, risk) that debate and refine a thesis before producing a final report.
Stateful LangGraph workflow with shared memory. Each specialist reads the running thesis, contributes a section, and writes back. A final arbiter agent reviews coherence and produces the deliverable. All steps are checkpointed for replay.
- 01Stateful LangGraph — every node is checkpointed; you can resume, replay, or branch the workflow at any step.
- 02Specialist isolation — each agent has a focused system prompt and a constrained tool set, which keeps hallucinations bounded.
- 03Local LLM workflows — runs against Ollama/vLLM with no API keys; designed for analysts who can't ship data to the cloud.