orchestratorCredit appraisal,
governed by design.
An AI orchestration layer for Vietnamese banks. Customer documents are read inside the bank. Only masked or public data goes out. A person makes every decision.
Runs entirely on
Automate the reading, keep the judgement. The framework decides what AI may do, where it may run, what evidence it leaves, and when a person must take over.
Analysts spend their judgement on data entry.
SME files mix scanned statements, registrations, bank records and valuations, all re-keyed by hand.
Extraction, OCR, research, memo drafting
Each task runs on the smallest model that clears the accuracy bar.
Deterministic skills and case memory
Units, separators, reconciliation and ratios are handled in code, not by the model.
Masking, egress scanner, review gate
Nothing identifying leaves. Every call is logged, and no memo is drafted before a person signs off.
From credit pack to decision in four steps
Read inside the perimeterSix documents are extracted in parallel by a sovereign 8B model.
Reconcile in codeCross-document checks flag revenue, charter capital, cash-flow and LTV issues.
Reason outside, maskedOnly public sector facts and a tokenised fact sheet reach the larger model.
A person decidesThe analyst disposes of every field, then the credit officer records the decision.
Synthetic data only. Masking recall on held-out adversarial formats was 73%, which is why raw documents stay inside the perimeter. Full method and limitations are in Progress Report 3.
An AI orchestration framework for document-intensive banking
Advisor: Dr. Nguyen Thi Mai Lan · College of Business and Management, VinUniversity