Aegis

Aegis: Dual-Engine AI for Resilient Risk Detection End-to-end research prototype for adaptive transaction risk scoring & explainability

Goal Detect anomalous/ fraud-like behavior in transaction streams Produce interpretable risk insights (RAG narrative) Measure model resilience under worst-case perturbations

1. Core idea Aegis builds a hybrid risk detection engine using: Sequence modeling (temporal Transformer) Graph representation learning (GNN on user-device-merchant graph) Fusion head to combine strengths Retrieval-Augmented Generation (RAG) for human-readable explanation Variance-based resilience testing on model scores

2. Data layer (data) Synthetic transaction simulation (synthetic_generator.py) user histories, timestamps, amounts, categories, geo plant risk events via known “fraud narrative” rules Preprocessing (preprocess.py) split train/val/test by user compute embedding-ready features, chronological sequences Output: sample CSV + risk narrative JSON graph edges for GNN sequence datasets for Transformer

3. Modeling layer Transformer path (transformer.py) per-user transaction sequences (time+amount+text features) temporal attention and classification head outputs per-transaction risk score GNN path (gnn.py) constructs heterogeneous graph: user ↔ device, user ↔ merchant uses GraphSAGE / GAT user-level embedding + risk score can capture cross-entity propagation patterns Fusion path (fusion_model.py) takes Transformer and GNN score vectors MLP fusion to final risk output learns weighted interaction + correction

4. RAG explanation (rag) Retriever (retriever.py) FAISS + sentence-transformers semantic search over risk_narratives.json Generator (generator.py) transformers text-generation model (distilgpt2 or OpenAI) condition on retrieved narratives + transaction context Pipeline (rag_pipeline.py) end-to-end narrative for a selected sample measures latency vs config threshold

5. Resilience testing (resilience_test.py) Stress variants: amount inflation/deflation time shifts merchant/device swaps For each sample, compute: score variance + conservative resilience factor 1/(1+var) Visual results: per-sample robustness aggregated stability plot

6. Evaluation (metrics.py) AUC-ROC for: Transformer-only GNN-only Fused model Other metrics: precision, recall, aggregated risk calibration In full pipeline (main.py): print comparative AUC with logs

7. Execution semantics (main.py) Generate and save dataset (+ narrative library). Fit transformer model. Build graph + fit GNN. Train fusion module. Evaluate with held-out users. RAG one sample + report latency. Run resilience evaluations and save plot. Store checkpoint for UI/re-use.

8. Produced artifacts outputs/sample_transactions.csv outputs/risk_narratives.json outputs/checkpoint.pt outputs/resilience_plot.png console metrics + explanation text