What is Sentinel?
SAGRA (Sentinel Adaptive Graph Risk Algorithm) is a graph-native adaptive fraud scoring engine powering Quantora AI. It combines transactional, topological, and network density signals into a single deterministic risk score for every transaction in real time.
Unlike traditional rule-based or row-level ML systems, SAGRA analyzes the structural context of each transaction within the broader financial network — detecting coordinated fraud rings, fan-out attacks, and layering schemes that single-transaction models miss entirely.
Research Lineage
Prototyped & validated in "SAGRA — Sentinel Adaptive Graph Risk Algorithm.ipynb" → Migrated to production backend (FastAPI + NetworkX).
Risk Formula
SAGRA computes four components per transaction, aggregated via weighted combination.
TRS = min(amount / 10000, 1)GRS = min(sender_degree × 5, 1)NDB = 0.3 if sender_degree > 0.05 else 0FRS = 0.5 × TRS + 0.3 × GRS + 0.2 × NDBFRS > 0.7 → FRAUD|FRS ≤ 0.7 → SAFERisk Components
Transaction Risk (TRS)
Normalizes transaction amount to [0,1]. Amounts approaching $10,000 carry maximum monetary risk due to regulatory reporting thresholds (BSA/AML). This is the strongest individual predictor, weighted at 50%.
Graph Risk (GRS)
Converts sender degree centrality into a risk signal. High-degree senders are statistically more likely to be involved in fan-out attacks, money mule networks, or layering schemes. Amplified by 5× to flag hub nodes early.
Network Density Boost (NDB)
Activates a +0.3 risk boost when sender connectivity exceeds a minimum threshold (degree > 0.05). This makes SAGRA adaptive — nodes becoming increasingly connected receive automatic risk escalation, catching 'smurfing' patterns.
Adaptive Thresholding
The NDB threshold (0.05) can be dynamically tuned per customer segment, time-of-day, or via online learning feedback loops. The decision boundary (0.7) and component weights can be optimized via grid search or Bayesian optimization.
Pipeline Architecture