Financial Statement Analysis Engine

Problem Being Solved

After a financial statement is uploaded, engineers and analysts often face inconsistent three-statement mappings, mixed account naming, and risk signals scattered across ratios and cross-checks. This skill parses .xlsx, .pdf, images, .docx, or pasted text into structured data, then runs Quick Diagnosis or Deep Diagnosis paths to produce risk conclusions.

How It Works

The pipeline routes files through parse_router.py: Excel/CSV extraction uses openpyxl and pandas, PDF parsing uses pdfplumber, images go through OCR, and Word tables are extracted. It then performs data validation, account mapping, and parallel computations, generating Z/M/F-Score, 21 red-flag checks, cash-flow patterns, and DuPont analysis. The three-tier funnel covers L1 health lights and a one-line diagnosis; L2 radar chart, industry percentile, risk heatmap, and three scenarios; L3 eight cross-checks, fraud-risk review, related-party penetration, and an action plan. Reports are HTML with Chart.js visualizations.

Boundaries and Notes

For Chinese .xls console garbling, follow the encoding-recovery checklist instead of guessing company or account names. Labels such as “abnormal” or “suspicious” indicate metric deviations, not legal conclusions. Data is processed in memory, and LLM calls are auto-redacted by removing entity names and range-ifying exact amounts. It can be deployed privately, but material decisions should still be reviewed by humans.