AML_Risk_Assessment_Demo_Presentation

1.

INTERNSHIP PROJECT
AI-Powered Anti-Money
Laundering Risk Assessment
TRANSACTION SIGNALS
for Financial Institutions Using Explainable Machine Learning
A transaction-level risk scoring prototype with model explanations and
assessment history.
EXPLAINABLE RISK
Presenter: [Your Name]
23 October 2026

2.

01 / CONTEXT
The challenge: prioritize suspicious activity
The goal is not to automate a compliance decision; it is to help analysts focus their attention.
Monitoring challenge
Project objective
What this prototype adds
Transaction monitoring can generate alerts
that require human review. Patterns may
involve the amount, timing, cross-border
context, and activity history.
Estimate transaction-level risk using
historical behaviour and transaction
characteristics, then provide an explanation
that an analyst can inspect.
Single-transaction assessment, batch CSV
assessment, SHAP contributions, a
dashboard, and saved assessment history.
DESIGN PRINCIPLE
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
Risk score + explanation + analyst review
D E M O S E S SI O N
2

3.

02 / SOLUTION
From transaction data to an explainable alert
A single assessment is scored, explained, and saved through the backend API.
INPUT
FEATURES
XGBOOST
SHAP
REVIEW +
HISTORY
Single transaction
or CSV batch
Transaction context
+ prior behaviour
Risk score and
threshold decision
Which features push
score up or down
Analyst investigates;
assessment is stored
Single assessment
Batch workflow
Score + alert decision + feature-level explanation
CSV in → per-row results out → saved assessment history
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
D E M O S E S SI O N
3

4.

03 / DATA
A time-aware evaluation on synthetic transaction data
The evaluation is designed to respect time order rather than randomly mixing past and future transactions.
500,000
11
129,471
2.19%
transaction research sample
model input features
held-out test transactions
positive share in test sample
Dataset construction
Temporal split
SAML-D synthetic transactions. All suspicious records were retained and
normal transactions were sampled by month to form the 500k research
sample.
Training: through 31 May 2023
Held-out test: 1 June – 23 August 2023
Threshold selection used a separate May 2023 validation window.
Important: the 2.19% test-set share is from the sampled dataset, not real-world AML prevalence.
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
D E M O S E S SI O N
4

5.

04 / FEATURES
Use behaviour, not raw identity, as model input
Account identifiers help build history features but are not passed directly into the predictor.
Transaction context
Sender history
Receiver history
• Amount
• Hour and day of week
• Cross-border indicator
• Currency-mismatch indicator
• Previous transaction count
• Previous total amount
• Number of unique prior receivers
• Previous transaction count
• Previous total amount
• Number of unique prior senders
LEAKAGE CONTROL
Historical features use earlier timestamps only; rows with the same timestamp do not see one another.
Raw IDs, timestamps, and laundering-type labels are excluded from model predictors.
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
D E M O S E S SI O N
5

6.

05 / MODELS
XGBoost performed best on the held-out test sample
All three models were evaluated on the same temporal test window and feature set.
METRIC
LOGISTIC REGRESSION
RANDOM FOREST
XGBOOST
Average Precision
29.46%
95.57%
97.20%
ROC-AUC
70.34%
99.46%
99.76%
Precision
54.00%
99.52%
99.22%
Recall
24.99%
88.19%
89.35%
F1 score
34.17%
93.52%
94.03%
False-positive rate
0.4770%
0.0095%
0.0158%
Model scores are from the sampled synthetic test set; real-world generalization remains unproven.
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
D E M O S E S SI O N
6

7.

06 / INTERPRETATION
Good headline metrics still need context
A confusion matrix shows both the alerts raised and the suspicious examples missed.
99.22%
89.35%
20
302
Precision
Recall
False positives
False negatives
H ELD - OU T T EST C O N FU S IO N M A T R IX
Actual label
Normal
Predicted normal
Predicted suspicious
TN 126,614
FP 20
FN 302
TP 2,535
Suspicious
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
How to read this responsibly
• 2,535 of 2,837 labelled suspicious cases were detected.
• 302 labelled suspicious cases were missed at this threshold.
• Alert decisions depend on the selected threshold and data distribution.
D E M O S E S SI O N
7

8.

07 / EXPLAINABILITY
SHAP helps explain what moved the model output
Global SHAP shows influential features across a sample; local SHAP explains an individual assessment.
TOP FEATURES BY MEAN ABSOLUTE SHAP VALUE
Global explanation
Sender: unique prior receivers
2.363
Receiver: unique prior senders
1.805
Receiver: prior transaction count
1.292
The features with the largest average absolute contributions
tended to describe prior sender/receiver behaviour.
Individual explanation
Receiver: prior total amount
0.987
Sender: prior transaction count
0.937
For one transaction, SHAP values show which features push
the model output toward higher or lower risk.
SHAP values are log-odds contributions — not percentage-point changes and not causal effects.
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
D E M O S E S SI O N
8

9.

08 / ENGINEERING
A separated frontend, API, model service, and database
The interface talks to the backend over HTTP; it does not load the ML model or access the database directly.
STREAMLIT UI
FASTAPI
ML SERVICE
DATABASE
Streamlit Community Cloud
Render backend service
XGBoost + SHAP
Supabase PostgreSQL
Dashboard • single • batch • history
/explain • /predict/batch • /history
Score • threshold • explanation
Persistent assessment history
API access control
Data persistence
Protected endpoints require an API key sent in a request header. Secrets are
configured in deployment environment/secrets, not committed to code.
Supabase PostgreSQL keeps production assessment history across app restarts.
SQLite is the local development fallback.
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
D E M O S E S SI O N
9

10.

09 / LIVE DEMO
A simple 3-minute walkthrough
Show one complete user journey instead of trying to demonstrate every feature.
01
Overview dashboard
Show system status and the main assessment summary.
02
Single transaction
Submit a transaction; show risk score, alert decision, and positive/negative SHAP contributors.
03
Batch CSV
Upload the prepared demo CSV; show row-level results and any validation feedback.
04
Assessment history
Refresh history and point out persistence of saved assessments.
30 sec
60 sec
45 sec
30 sec
Backup: preload the app, keep the demo CSV locally, and have one screenshot of a completed assessment ready.
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
D E M O S E S SI O N
10

11.

10 / CONCLUSION
A useful prototype — with clear next steps
The contribution is the end-to-end workflow and explainable risk assessment, not a claim of production readiness.
What is delivered
What remains to validate
• Three trained and compared models
• Time-based evaluation and fixed threshold
• SHAP global and local explanations
• Deployed UI, API, batch workflow, and persistent history
• Synthetic data limits real-world generalization
• Sampled class balance is not natural prevalence
• Thresholds need institution-specific calibration
• Fairness, drift, security, and analyst workflow need further testing
CORE TAKEAWAY
Machine learning prioritizes. SHAP explains. A qualified analyst decides.
Thank you • Questions?
A M L R I S K A S S ESSMENT / I NT E R NSHI P P R O J E CT
D E M O S E S SI O N
11