THE PROBLEM
Why this system existed
Transaction review spans heterogeneous documents and requires consistent findings, cross-document checks, and senior approval.
Agents · Document Intelligence · Risk
Deal Advisory Due Diligence Agent
A six-step agentic pipeline from document classification through extraction, entity resolution, risk scoring, consistency checks, and structured deal-memo generation.
MY ROLELangGraph orchestration, document extraction, risk scoring, and review controls
SCOPEA multi-document workflow that classifies financial and transaction documents, extracts structured findings, checks consistency, and produces reviewer-ready summaries.
EVIDENCEProfessional experience · public-safe summary
5 → 1.5business days
14pilot engagements
23discrepancies surfaced
OUTCOME
What changed
Reduced first-pass due-diligence summary turnaround from 5 business days to 1.5 days across 14 pilot engagements.
REFERENCE ARCHITECTURE
Controls around the model
01Classify
02Route
03Retrieve
04Rerank
05Generate
06Validate
07Observe
DECISIONS
Trade-offs considered
Explicit six-step state graph
Structured risk severity
Cross-document consistency checks
Partner-level review gate
FAILURE CASE
What did not work
A fluent summary can hide an inconsistency. The workflow therefore checks entities and claims across documents before generating the final structure.
SECURITY BOUNDARY
Public-safe by design
This case study exposes patterns, not employer architecture. It uses synthetic data, no client identifiers, no internal prompts, no proprietary datasets, and no production endpoints.