Take-private Credit Analysis Agents
A ReAct agent pipeline that drafts credit committee materials from deal documents, with every number cited or computed.
refining
Summary
Takes a CIM, credit agreement, audited financials, QoE, and sponsor model, and produces the pieces of a sponsor-financing credit memo: EBITDA reconciliation, sources and uses, pro forma cap table, returns and pricing analysis, risk write-ups, and company and industry overviews, assembled into a PowerPoint deck. A review agent runs a consistency and quality pass before a human sees it.
Problem
Credit committee prep is days of extracting numbers from long PDFs and Excel models, retyping them into templates, and cross-checking that everything ties. Most of that is mechanical.
AI Usage
LLMs extract and write; code calculates. Every number in the output is either extracted by an agent with a citation (doc and page, or workbook cell) or computed by deterministic Python from a canonical DealWorkspace schema. The LLM never freehands the EBITDA bridge, cap table, or returns math. A review agent runs deterministic cross-checks plus a judgment pass. Only sanitized or made-up deals go in until a compliance-approved endpoint exists.
- Every extracted figure carries a citation, so a reviewer can trace it to the source page or workbook cell.
- Deterministic engines for EBITDA bridge, sources and uses, cap table, case adjustments, and IRR/MOIC/YTM.
- The deck's first content slide is always QC status, so nothing looks finished while open issues remain.
- Model provider centralized in one setting, so moving to an enterprise endpoint is config, not a rewrite.
Stack
Python, Anthropic SDK, Pydantic, python-pptx, PDF and Excel tooling