Blog
Sovereign compliance, in depth.
Practical writing on the architecture, the regulations and the modeling behind KYC, AML, fraud, credit scoring and screening — where the data never leaves.
Why your customer data should never leave your infrastructure
The regulatory, security and commercial case for keeping raw data on-premise — and how near-data compute makes it practical without sacrificing AI.
KYC onboarding: from manual review to risk-based automation
How to move customer due diligence from queue-clearing to a risk-based model that scores every applicant in seconds — without exporting identity data.
AML transaction monitoring without moving your data to the cloud
Transaction monitoring generates some of the most sensitive data a bank holds. Here's how to score it for money-laundering risk without exporting a single row.
Detecting structuring and smurfing with typology scenarios
Structuring is designed to stay under reporting thresholds. Here's how aggregate windows and network signals catch it — scored on your own data.
Real-time payment fraud scoring at sub-100ms
Fraud decisions must happen before the payment clears. How serving a model in-memory near the data delivers low-latency scoring without shipping transactions out.
Account takeover detection: the signals that actually matter
Account takeover looks like legitimate activity from a legitimate user. The behavioral and device signals that separate the real customer from the impostor.
Explainable credit scoring with SHAP, built for regulators
A credit score a regulator can't interrogate is a liability. Here's how SHAP turns model predictions into defensible, per-decision explanations.
IFRS 9 staging and expected credit loss, explained
A practical walkthrough of IFRS 9 stages 1, 2, 3 and POCI, and how to compute expected credit loss provisioning on your own data — auditable and reproducible.
Sanctions and PEP screening: cutting false positives with fuzzy matching
Name screening drowns teams in false positives. How fuzzy matching on aliases, dates of birth and countries — plus a review queue — makes screening usable.
How AI agents write compliance code without ever seeing your data
Two AI agents build your use cases — but they reason over metadata, not raw rows. Here's exactly where the confidentiality boundary sits.
Near-data computing: bring the compute to the data
The cheapest, fastest and safest place to run a computation is next to the data. Here's why near-data compute is the right default for regulated workloads.
Outbound-only runners: the GitHub Actions model for compliance
Why a data plane that pulls work over an outbound connection — and never opens an inbound port — is the safest way to connect on-premise compute to a SaaS.
GDPR and data residency for EU banks: an architectural approach
Data residency is easier to prove than to promise. Here's how a data-plane architecture turns a compliance headache into an architectural fact.
Model risk management and MLOps for compliance teams
SR 11-7 and EBA expectations meet modern MLOps. What model governance looks like when validation, drift monitoring and documentation are built in.
Deploy a compliance data plane in 15 minutes
Self-serve onboarding, step by step: from sign-up to first scores in about fifteen minutes, with a single command to install the Runner.
Choosing an isolation tier: standard, restricted, or 100% local
Not every institution can let the same data leave. KeplerAI's isolation tiers match data movement to your regulatory posture — up to nothing leaving at all.
Entity resolution and network graphs for fraud rings
Fraud rings hide in the links between accounts. Here's how entity resolution and relationship graphs expose coordination that per-account rules miss.
Data quality profiling: the foundation of trustworthy models
Every compliance model rests on the quality of its inputs. Here's what to profile — and why doing it near the data matters.
Unified compliance vs point solutions: the real total cost
One vendor for screening, another for fraud, another for scoring. The license fees are only the beginning of what fragmentation costs.