A bank’s data grows every day, and along with it the demand on the data layer: regulatory reports, fraud-detection models, customer-360 dashboards, and over the last two years a wave of AI assistants. When the data team wakes up to a "why is this report different today?" question, they often cannot tell who in the stack to trust. CentraQL measures, explains and protects that trust.
CentraQL is not a rule engine
The classical DQ approach is: write a library of SQL rules, run them overnight, paint the dashboard red/green. CentraQL bundles a wider surface into a single platform: catalog, contract-driven data quality, KPI anomaly detection, lineage, audit reporting — plus a natural-language AI Copilot on top.
What does it solve?
- Pipeline reliability: trust score, contract violations, freshness and anomalies are tracked on the same dashboard.
- Decision time: "top-5 most profitable branches last month" is answered by a semantic-layer-grounded Copilot with verified SQL and a narration — no ticket to IT.
- Regulatory alignment: PromptAuditLog, ColumnPolicy masking and the RegulatedFinance profile let you close the query surface to BDDK, KVKK and EU AI Act requirements.
Architecture
CentraQL is a modular monolith on .NET 9. Modules: Catalog, DataQuality, KpiAnomaly, SemanticLayer, Embeddings, Copilot, Insights, AuditAnalytics, Onboarding, and Worker. SQL Server underneath, Qdrant for vectors, Redis for cache. The LLM layer can run fully on-prem (Ollama, vLLM, TEI) and Anthropic Cloud is available behind EgressGuard for those who want it.
The AI BI Copilot, briefly
A natural-language question enters an 11-stage governed pipeline that is fully audited: Guard → Intent → Retrieve (Qdrant) → Plan (LLM, QuerySpec JSON) → Validate → Synthesize (SQL) → Execute (read-only) → Mask (PII) → Chart → Narrate (LLM) → Audit. The answer always passes through the semantic layer and the catalog boundary; number, currency and total sanitization runs at the end.
Why does this matter for a bank?
Banks do not need just a BI dashboard — they need a provable answer. The layered discipline of CentraQL (Trust Score + PromptAuditLog + ColumnPolicy + ComplianceProfile) makes the answer regulator-grade traceable. On top of that, an analyst or a CRO can ask in natural language and get a response within minutes, with no IT bottleneck.
How do you start?
A typical start runs through:
- Discover — schema discovery and sample profiling on the connected sources.
- Contracts + Rules — first wave on 5-10 critical tables, contract + rule definitions, suggestions inflate the rest.
- KPI definitions + anomaly thresholds — 10-15 KPIs to start, threshold + z-score together.
- Copilot pilot — banking-tr or banking-en domain pack, a handful of certified few-shots, a focused two-week pilot.
- Audit & compliance switch-on — RegulatedFinance profile, EgressGuard, retention policies.
For a CentraQL live demo please get in touch (https://bi-art.com.tr/en/contact) — we set up the environment with bank-shaped demo data tailored to your scenario. This is the first article in the category; the next ones will dig into the Copilot pipeline, the Trust Score formula, the anomaly models and bank-specific learning via LoRA.
