↑ back to the story

enterprise case study · anonymized

A document-understanding platform for a regulated Malaysian bank

How I designed, deployed, and operationalized an AI document-understanding platform inside a regulated, data-resident, human-in-the-loop environment, where “move fast and break things” is not an option.

“In a regulated bank, the model is the easy part. The hard part is everything that has to be true around it: residency, auditability, and a human who can always say no.”
client
A regulated Malaysian bank, named under NDA on request
my role
Lead solutions architect & delivery engineer: architecture, build, deployment, handover
environment
Data-resident, on-prem / private-VPC · human-in-the-loop by design
disclosure
High-level architecture, timeline & qualitative outcomes only
status
delivered · in production
01

the architecture

bank-controlled perimeter · on-prem / private vpc · residency · least-privilege
1
secure ingestion
2
OCR & normalize
3
layout & table parse
4
LLM extract + confidence
5
validation & rules
6
structured output
↳ low-confidence → human review (accountable)
═ audit & observability spine · lineage · model version · confidence · human action · outcome ═
02

engineering decisions & tradeoffs

In-perimeter model serving over hosted APIs

Residency was non-negotiable, so no public multi-tenant API. More infra to operate, and in return you get control, cost predictability, no per-token vendor lock.

Confidence-gated human review over full automation

Full automation demos better; it's indefensible here. Thresholds route only uncertain/high-risk items to a person: the human is the control.

Modular swappable stages over a monolith

Each stage validated, versioned, upgraded independently: an engineering win and a compliance property.

Audit as a first-class spine over logging bolted on

Retrofitting auditability is how regulated projects fail their first governance review. Built in from day one.

03

outcomes

  • A production platform, live inside a regulated bank: Past the governance bar that stops most bank-AI at the prototype stage.
  • Staff moved from transcription to judgment: People review only what genuinely needs a human.
  • Audit-ready by construction: “What happened and who signed off” is a query, not a scramble.
  • A capability the bank owns: Delivered with runbooks and knowledge transfer, on a modular architecture they can extend.

disclosure & verification

Intentionally anonymized: I never publish the client's name, their data, or invented metrics. Client name, engagement specifics, and figures are available to serious counterparties under NDA.