SENTRIX AI — Fraud Intelligence Platform
Detect. Decide. Defend.
SENTRIX AI is a fraud intelligence platform concept for card and UPI transaction risk assessment. It combines engineered transaction features, an XGBoost baseline, SENTRIX-LM (a domain-adapted language model), and a policy-based decision engine to support explainable fraud decisions. The system explores how predictive modeling, policy-based decisions, and carefully bounded language-model capabilities can work together in a transaction-risk workflow.
≥ 85%
Fraud recall target
≥ 0.80
PR-AUC target
≤ 250 ms
p99 latency target
Figures labelled target or est. are design targets or estimates, not measured results.
20-week delivery plan
Weeks 1–2
Data quality, governance, and architecture design.
Weeks 3–5
Feature pipeline and XGBoost baseline.
Weeks 6–8
Model shortlist, pilot tuning, go/no-go decision.
Weeks 9–13
Model development, optimization, scoring API.
Weeks 14–16
AWS deployment, security, CI/CD, observability.
Weeks 17–19
Shadow pilot, acceptance tests, operational readiness.
Week 20
Documentation, runbooks, and knowledge transfer.
Problem & use case
Legacy rule-based fraud engines generated high false-positive rates, lacked contextual understanding of novel attack patterns, and produced black-box decisions that fraud investigators and compliance teams could not interpret or defend to regulators.
Solution & engineering approach
A hybrid architecture where XGBoost provides the mandatory tabular baseline and SENTRIX-LM supplies domain-adapted language-model scoring. Both feed a Policy & Decision Engine that applies ALLOW / REVIEW / BLOCK thresholds and hard-rule overrides — with explainable rationale generated for every decision.
Built around three principles
01
Engineering-Led Intelligence
The LLM is not automatically the best fraud scorer. SENTRIX-LM must prove its value against the XGBoost baseline. If it fails to achieve the required uplift, the approved fallback is a hybrid: XGBoost scores, LLM explains.
02
Domain-Adapted Language Model
SENTRIX-LM undergoes continued pretraining on fraud corpora, supervised fine-tuning on labeled transactions, and preference tuning — shortlisted from open-weight models in the 1B–8B parameter range.
03
Private, Secure Serving Infrastructure
Designed for private cloud deployment — private EKS, GPU inference nodes, Redis velocity store, mTLS, service-token auth, Terraform IaC, and zero reliance on third-party LLM APIs. Serving infrastructure targets low-latency processing and operational resilience.
Architecture & data flow
- 01
Transaction input: card / UPI event and customer context ingested via REST API
- 02
Feature engineering: velocity, device, location, account age, merchant and behavioral signals
- 03
XGBoost baseline scores fraud probability — mandatory benchmark for all LLM comparisons
- 04
SENTRIX-LM (1B–8B open-weight model) produces domain-specific fraud signal and explanation
- 05
Policy & Decision Engine applies thresholds and hard rules → ALLOW / REVIEW / BLOCK verdict
Technology stack
Evaluation methodology
- Evaluation: decision traceability, explanation quality, and operational resilience by design
Design targets — not measured results
- SENTRIX-LM must prove ≥ +0.03 PR-AUC uplift over XGBoost baseline (design target — not assumed, earned)
- Precision at BLOCK ≥ 70% · REVIEW queue ≤ 12,000/day (design targets)
- Explainability: ≥ 90% factual correctness of fraud decision rationale (design target)
- Fraud losses ≤ 6.4 basis points across the portfolio (design target)
Implementation status
In Development — architecture, design, and model approach defined; implementation ongoing
A Saffronyx AI project — Designing Intelligent Systems That Last.
SENTRIX explores an engineering-led approach to fraud intelligence in which predictive models, decision policies, operational controls, and evaluation are treated as parts of one system. Numerical results will be published only when supported by documented testing; any proposed thresholds are targets rather than achieved outcomes.
Bring the idea. We forge the system. — Saffronyx.in