Explainable AI Agents for Transparent Financial Decision-Making

A production-ready multi-agent Explainable AI system for transparent, auditable financial decision-making. An Orchestrator Agent coordinates a pipeline of specialized agents to generate predictions, feature attributions, and human-readable narratives in a single traceable process.

---

Table of Contents

• Architecture
• Results
• Repository Structure
• Quick Start
• XAI Method Selection
• Testing
• Visualizations
• License

---

Architecture

Agent	Role	Key Function
Orchestrator	Workflow Control	Manages end-to-end process: data collection, decision, XAI, narrative, audit logging
Decision Agent	Prediction	Executes the trained financial model to produce a prediction
XAI Agent	Attribution	Selects and runs the appropriate XAI method (SHAP, LIME, etc.)
Explanation Agent	Narrative Generation	Translates XAI attributions into human-readable, context-aware narratives
Evidence Collector	Context and Metrics	Gathers model performance metrics and contextual data for the audit log
PII Redactor	Privacy	Redacts sensitive PII before processing or logging

---

Results

Model Performance

Model	ROC-AUC	Accuracy	F1 Score	Inference Time
Random Forest	0.732	0.715	0.728	3.8ms
Logistic Regression	0.685	0.692	0.710	2.1ms
Neural Network	0.698	0.701	0.715	4.2ms
Full XAI System	0.732	0.715	0.728	340ms (with explanation)

XAI Method Comparison

Method	Avg Time (ms)	Memory (MB)	Faithfulness	Best Use Case
SHAP	850	450	0.85	Regulatory compliance, high-stakes decisions
LIME	180	80	0.75	Real-time systems, large-scale deployment
Integrated Gradients	320	200	0.80	Neural networks, research
Counterfactual	2300	60	0.70	What-if analysis, customer-facing

Human Trust Study

System	Trust Score (1-5)	vs Baseline
Black-box Model	2.9	baseline
Model + SHAP	3.6	+24%
Model + LIME	3.4	+17%
Full XAI Agent System	4.1	+41%

---

Repository Structure

Path	Description
code/agents/	Orchestrator, XAI agent, explanation agent, privacy redactor
code/models/	Baseline model definitions and comparison logic
code/xai/	XAI method implementations and selection logic
api/	FastAPI server for predictions and explanations
scripts/	Demo runner, experiments, visualization suite
deployment/	Docker Compose and Kubernetes configuration
monitoring/	Prometheus and Grafana configuration
visualizations/	8 generated publication-quality figures
docs/	XAI method selection guide and documentation

---

Quick Start

Comprehensive Demo

git clone https://github.com/quantsingularity/Explainable-AI-Agents-for-Transparent-Financial-Decision-Making.git
cd Explainable-AI-Agents-for-Transparent-Financial-Decision-Making

pip install -r requirements.txt -r requirements-api.txt

python code/scripts/comprehensive_demo.py

Outputs model and XAI performance metrics to console and saves 8 figures to visualizations/.

Production API

# Local
python api/api_server.py

# Docker
docker-compose -f deployment/docker/docker-compose.prod.yml up -d

Endpoint	Method	Description
/health	GET	Service health check
/predict	POST	Financial prediction via Decision Agent
/explain	POST	Full XAI explanation and narrative
/metrics	GET	Prometheus metrics endpoint

API docs available at http://localhost:8000/docs.

---

Output

Console output stays readable: benign third-party warnings and library progress bars are suppressed, so only meaningful log lines remain. Each run finishes with a clean, aligned summary block reporting a per-configuration table with the best AUC and most faithful XAI method highlighted.

XAI Method Selection

Requirement	Recommended Method	Reasoning
Regulatory compliance	SHAP	Highest faithfulness and theoretical soundness
Real-time scoring under 200ms	LIME / Feature Importance	Fastest, model-agnostic
Tree-based models (RF, XGBoost)	SHAP TreeExplainer	100x faster than Kernel SHAP
Neural networks	Integrated Gradients	Best for gradient-based non-linear models
What-if analysis	Counterfactual	Actionable insights for outcome changes
Large datasets over 10k samples	LIME	More scalable than SHAP KernelExplainer

Full decision tree and best practices: docs/XAI_METHOD_SELECTION_GUIDE.md

---

Testing

pytest code/tests/ -v --cov=code --cov-report=html

Type	Coverage	Location
Unit Tests	82%+	code/tests/
Integration Tests	Full pipeline	code/tests/test_baseline_and_xai.py
Performance Tests	Latency tracking	code/xai/xai_methods.py

---

Visualizations

Running visualization_suite.py generates 8 publication-quality figures saved to visualizations/.

Figure	Title
01	System Architecture
02	XAI Method Comparison
03	Feature Importance
04	Model Performance
05	Human Trust Results
06	XAI Performance Trade-offs
07	Latency Breakdown
08	XAI Decision Tree

---

License

Licensed under the MIT License. See LICENSE for details.