Capstone: end-to-end AML monitoring proposal for a mid-tier Pakistani bank

Three ways to see it

1. 1

   A defensible 12-month plan has six workstreams. Workstream one is governance: build the model inventory, charter the model risk committee, define the champion-challenger protocol, and contract an independent validator for the year-end review. This is the cheapest workstream, the highest in regulatory leverage, and the one banks most often defer. Workstream two is feature engineering: stand up a feature store, build the Raast-specific feature pack covering velocity, structuring, counterparty diversity, geography entropy, and turnover-to-declared ratio, and run a feature governance review on every existing rule input. Workstream three is the ML monitoring layer itself: a supervised XGBoost model for known typologies, an unsupervised isolation forest for novel patterns, deployed in shadow mode for 90 days before any score actions a hold.

2. 2

   Workstream four is graph capability: stand up Neo4j or TigerGraph, model accounts as nodes and Raast plus IBFT transactions as edges, compute degree, PageRank centrality, and connected-component features as a daily batch, and feed those features into the supervised model from workstream three. This single change has historically produced the largest single jump in mule-ring detection in similar banks. Workstream five is sanctions screening: replace any vendor matcher that does not have a Pakistan transliteration pack, set per-list thresholds with documented OFAC miss-cost rationale, and verify same-business-day re-screening capability against new list entries. Workstream six is investigator workflow: enable SHAP local explanations on every alert, redesign the case management screen around the six-element STR template, and rotate analysts off triage every two hours.

3. 3

   Sequencing matters. The naive plan is to start with the ML model because it sounds the most exciting. The defensible plan starts with governance and feature engineering, because an ML model deployed without governance is a Matter Requiring Attention waiting to be discovered, and an ML model trained on ungoverned features inherits whatever bias the features encode. A practical 12-month sequence is: months 1 to 3, governance and feature store; months 4 to 6, supervised model in shadow mode plus sanctions matcher upgrade; months 7 to 9, graph capability online and feeding the supervised model, ML layer promoted to action with a high threshold; months 10 to 12, isolation forest and explainability rolled out, threshold tuning, independent validation, Board sign-off. Slipping any earlier step into a later month is fine; reordering them is not.

The why-tree

Try this with Claude

AI-edge prompt: 'I am the Head of Financial Crime Technology at Saraab Bank, a fictional Pakistani mid-tier bank with the profile in this lesson. Draft me the full Board memo for the 12-month AML transformation programme. Include executive summary, the six workstreams, the 12-month sequence, the budget table in PKR, the four named risks with mitigations, the success metrics broken down per quarter, and the questions I should expect the Board Risk Committee to ask. Write it in plain English suitable for non-technical Board members; assume the Board includes one ex-SBP deputy governor.' Use the model output as a starting draft, not the final memo.