FinShield Underwriting RAG
Project Overview
FinShield Underwriting RAG is an enterprise commercial credit, KYB, and financial covenant analysis platform built by Eternitech for a B2B specialty lender processing thousands of middle-market credit applications per quarter.
The Architectural Challenge
Each underwriting package combined 200-to-450-page scanned PDF tax returns, multi-tab debt schedules, Plaid cash-flow feeds, and Secretary of State filings. Early vector-only RAG prototypes built internally by the client misread multi-column balance sheet tables and blended figures across fiscal years—making credit committee sign-off impossible.
Engineering Solution & Tech Stack
Eternitech replaced naive fixed-token chunking with a deterministic, table-aware financial ingestion and hybrid retrieval architecture backed by PostgreSQL pgvector and row-level tenant security.
- Layout-Aware Table & Ledger Parsing: Document layout models reconstruct balance sheets, P&L statements, and debt-service schedules into canonical Markdown + typed JSON rows before embedding, preserving fiscal-year column headers on every chunk.
- Hybrid Retrieval + Cross-Encoder Reranking: Reciprocal Rank Fusion (RRF) merges dense embeddings with PostgreSQL BM25 exact account-code matching, followed by cross-encoder reranking.
- Deterministic Python Ratio Engine: Rather than asking an LLM to compute DSCR, EBITDA adjustments, or cash-burn multiples in prose, tool-calling delegates all arithmetic to audited Python financial functions.
- Inline PDF Bounding-Box Citations: Every ratio and risk flag in the credit memo clicks directly to the exact highlighted cell in the source PDF viewer.
Quantified Production Outcomes
| Metric | Manual Underwriting | FinShield Production |
|---|---|---|
| End-to-End Credit Memo Prep Time | 4.2 business days | 38 minutes (-98% elapsed) |
| Numerical / Fiscal-Year Hallucinations | 14% in naive RAG prototype | 0.0% (Python-verified math) |
| Analyst Underwriting Throughput | 6 deals / week | 22 deals / week (3.6x) |
See how our RAG & Vector Search Engineers and Fintech Architecture Pods build audit-ready financial AI.
Want the same operator pod behind FinShield Underwriting RAG on your roadmap?
Explore our AI Agents & RAG Engineering, Dedicated Engineering Pods, or book a 30-minute scoping call with a Technical Project Manager.
Tell us what you're building.
No pitch deck required. Just a description of what you need. We'll come back with a scope breakdown — and we'll be honest if we're not the right fit.