FINANCE & EQUITY RESEARCH
Financial & research report multi-table analysis
Target teams: Equity research, corporate investment, and FP&A teams comparing 200–400 page annual reports and broker research PDFs.
Core challenge: Merged headers, multi-column financial statements, and cross-page footnotes get scrambled by naive text extraction, while direct LLM mental math produces unreliable YoY growth and CAGR figures.
Verified architecture: Layout-aware PDF parsing preserves table headers, currency units, and page mappings; an offline-sandbox Code Mode agent retrieves raw figures across multiple reports and runs Python scripts for unit normalization, cross-check reconciliation, charts, and Excel exports; reviewers verify every number in the sentence-level source drawer.
Acceptance & governance gate: Every extracted figure cites its source report and footnote; all ratio, YoY, and reconciliation calculations execute deterministically in Python rather than LLM mental arithmetic.
Key primitives
- Layout-aware PDF & complex table parsing
- Offline Code Mode Python sandbox
- Sentence-level source verification drawer
- Reusable org skills & read-only functions