Integrity Stack’s flagship product, Sentinel, intercepts every AI agent decision in real-time, evaluates it against behavioral rules and security guardrails, and returns a verdict—APPROVE, BLOCK, FLAG, or ESCALATE—before the action executes.
Every decision passes through a multi-layer defense pipeline before a verdict is returned. Each stage is independently configurable per agent type.
Seven pillars of AI governance, observability, agent trust, and control—built for regulated industries from day one.
Prompt injection scanner (7 threat categories including social engineering, multi-turn attacks, and homoglyph detection), 11-dimension safety scorer, toxicity filter, PII/PHI redactor (24 entity types), faithfulness/groundedness checker, rate limiter. All configurable per agent type via API.
Session/trace/span model for full agentic lifecycle visibility. Token and cost tracking across GPT-4o, Claude, and custom models. Performance metrics with p50/p95/p99 latency, throughput, and error rates.
10 regulatory frameworks pre-mapped (ECOA, FCRA, TILA, GDPR, RBI, SOX, PCI-DSS, BSA/AML). NIST AI RMF and EU AI Act alignment. SR 11-7 model inventory. Compliance reports with risk scoring.
Plain-English decision summaries, per-rule explanations for all 9 condition types, feature contribution scoring, counterfactual suggestions ("increase credit_score from 580 to 620"), and full decision path tracing.
KYA API and dashboard: dossier per agent (intended use, inferred data categories, template lineage, active rules, compliance tags). Agent explainability narrative, decision flow, and rule catalog. Proof of Agent: every evaluation returns SHA-256 bindings tying the verdict to registry state; audit logs store canonical replay strings; optional X-Proof-Of-Agent possession secret.
Bias monitoring with 4/5ths rule fairness checks, disparate impact tracking, adverse action reason codes (ECOA Regulation B), PSI-based drift detection, and data quality scoring across all inputs.
API key authentication with SHA-256 hashing, SSO/SAML/OIDC enterprise auth, role-based permissions, human-in-the-loop ESCALATE verdict with approval queue, alert rules engine with cooldown, and full audit trail export (JSON/CSV).
Every guardrail is independently configurable with tunable thresholds per agent type. No one-size-fits-all.
Detects and redacts 24 PII types including SSN, credit cards, Aadhaar, PAN, IBAN, passport, driver's license, medical records, API keys, and more. Replaces with [REDACTED_TYPE] before any processing.
Scores text across illegal, hateful, harassing, racist, sexist, violent, sexual, harmful, unethical, jailbreaking, and roleplaying. Per-dimension scores 0.0-1.0 with configurable thresholds.
Scores LLM responses against RAG context using token overlap, n-gram coverage, and claim verification with 5% numeric fuzzy matching. Catches fabricated numbers and hallucinated facts.
Detects 7 threat categories: system prompt overrides, role manipulation, data exfiltration, encoding attacks, delimiter injection, social engineering, and multi-turn attacks. Homoglyph normalization included. Immediate BLOCK with confidence scoring.
Pre-mapped compliance frameworks with automated reporting, adverse action codes, and audit evidence generation.
All 4 functions (GOVERN, MAP, MEASURE, MANAGE) mapped to Sentinel capabilities. 21 categories with specific coverage evidence.
4-tier risk classification with obligation tracking. 9 Sentinel controls mapped to HIGH risk requirements (data quality, logging, transparency, human oversight).
Model inventory with periodic review tracking. Auto-generates SR 11-7 compliance reports flagging models overdue for review.
Adverse action reason code generation (7 standard codes + default). Max 4 reasons per ECOA requirement. Prohibited factor detection.
Credit reporting compliance tags, Truth in Lending checks, Qualified Mortgage DTI limits (43%). All enforced at the rule level.
Data privacy controls, KYC/AML framework support, payment card industry standards. 24-type PII detection covers international identifiers.
The deepest regulatory knowledge base of any AI governance platform. Per-agent compliance requirements with examiner expectations and regulator citations.
"A creditor's lack of understanding of its own methods is not a cognizable defense." Sentinel ensures every AI lending decision is explainable and auditable.
"Banking decisions must be well-controlled, numerically and legally precise, explainable, and replicable." Sentinel's explainability module delivers exactly this.
National banks must apply existing risk management frameworks to AI. Sentinel maps to SR 11-7 and OCC guidance with automated compliance evidence.
GAO identified AI bias in lending as significant risk and found NCUA lacks AI vendor examination authority. Sentinel's bias monitoring fills this gap.
"Over-regulation of AI can itself present risks." Existing frameworks like SR 11-7 are sufficient. Sentinel ensures compliance without over-engineering.
Interagency rule requiring quality control for AI-powered automated valuation models in mortgage. Sentinel monitors AVM outputs for bias and drift.
Oliver Wyman's framework identifies 5 AI agents orchestrating the end-to-end credit chain, all requiring "codified risk appetite & policies" and "purposeful human steer." Their Phase 2 ("Codify") literally describes Sentinel: guardrails for AI risk management, risk-based escalation logic, codified institutional expertise.
Self-hosted, zero data egress. Deploy in your VPC or on-prem. No external API calls, no hidden costs.
3 lines of code to integrate. SentinelClient wraps all API calls with automatic retry and error handling.
SentinelGuardrail and SentinelCallbackHandler for seamless integration with LangChain agent pipelines.
88 endpoints with OpenAPI docs. Works with any language or framework. Webhook alerts for BLOCK/FLAG decisions.
One API call deploys a fully configured agent with compliance-tagged rules and BFSI-specific guardrails. Per-agent regulatory mapping from 9 U.S. regulators.
Amount limits, income ratios, term validation, escalation thresholds
Credit score range, DTI limits, employment history, FCRA compliance
Required fields, age verification, income validation, TILA/ECOA tags
Usury rate caps, prohibited factor detection, fair lending compliance
Bureau validation, credit history, inquiry limits, FCRA compliance
Valid FICO range, prohibited factors, fair lending enforcement
Protected class detection, disparate impact checks, ECOA/FHA
Income-based limits, minimum thresholds, CARD Act compliance
AVM integration, HMDA reporting, QM/ATR rules, FHA compliance
Dealer markup caps, MLA compliance, rate discrimination prevention
Co-signer evaluation, state licensing, ECOA/TILA compliance
UDAAP compliance, TCPA consent, pre-screened offer rules
CLI adverse action, CARD Act rules, income re-verification
Hybrid platform + usage pricing. Budget-predictable for bank CFOs. No surprise bills.
For evaluation & testing
For community banks & credit unions
For mid-size banks & fintechs
For large banks & insurers
All plans include the full 11-stage security pipeline. Annual billing: 2 months free. Overage: $0.03-$0.05/eval.
Deploy Sentinel by Integrity Stack in your environment in minutes. Self-hosted. Zero data egress. 1,659 tests. Production-ready.
Ask about Sentinel, compliance, or AI decision integrity