BullClaw

A multilingual financial report generation system built on LLM-based multi-agent workflows.

BullClaw is a multilingual financial report generation system designed for large-scale earnings analysis with LLM-based multi-agent workflows.

  • Built a stateful LangGraph workflow with specialized retrieval, extraction, writing, and compilation agents.
  • Supported automated analysis of financial reports from 8 major global exchanges in 4 languages, addressing long-context bottlenecks in LLM pipelines.
  • Introduced Sirchmunk (Agentic Search) as the core RAG engine and combined it with structured Markdown chunking to form an “extract, calculate, then explain” workflow.
  • Reduced numerical hallucinations in cross-lingual and cross-period financial reasoning.
  • Ranked 3rd overall on OpenFinArena / FinDocResearch, while placing 1st in the U.S. market and 2nd in Mainland China, outperforming strong baselines such as DeepSeek-v3.2.
  • Added CLI-based batch execution and checkpoint recovery for more robust large-scale generation.