Decision intelligence
Analytics that lights the decision, not just the dashboard.
Lumen holds one approved definition for every KPI and watches those metrics continuously. It detects what changed, explains why, forecasts what comes next and recommends what to do, with the expected impact stated. Ask in plain language: answers come only from governed metrics and show the query behind them.
VARDA LumenRuns on Orbit
The problem
Reporting what happened is no longer the hard part.
Every organisation can say what happened last month. Saying why it happened, what is likely next and what to do about it is harder — especially when teams still disagree about the number itself.
- 01
The same KPI is calculated one way in finance and another in sales.
- 02
Changes surface at month-end, long after they could have been acted on.
- 03
Explaining a variance takes days of manual slicing in spreadsheets.
- 04
AI chat tools answer confidently from data nobody has validated.
- 01
What happened?
Region B net sales fell 8.4% week on week.
- 02
Why did it happen?
71% of the drop comes from stock-outs in two product groups.
- 03
What is happening now?
Stock is below the critical level in 3 stores.
Store 14Store 22Store 31 - 04
What happens next?
Forecast for the next 4 weeks, with its uncertainty band:
- 05
And most importantly — what should we do about it?
Transfer 1,200 units from Warehouse 2 to Store 14
Expected impact: +₺186k net sales · range ±₺40k
Send for approvalShow the querySELECT net_sales FROM metrics.sales WHERE region = 'B' AND week = current_week()
Capabilities
Not just what happened. What to do next.
Semantic metrics layer
Each KPI is defined once — formula, grain, filters, currency and calendar — versioned like code and approved by its owner. Every dashboard, alert and answer reads the same definition.
Anomaly detection and alerting
Every metric is watched against its own seasonality and trend. Related anomalies are grouped into a single alert and sent to the metric’s owner, ranked by business impact rather than statistical score alone.
Root-cause decomposition
When a metric moves, Lumen breaks the change down by region, product, channel and customer, separating price, volume and mix effects. The real driver becomes visible without anyone slicing the data by hand.
Forecasts with uncertainty
Forecasts come with uncertainty bands, not a single line. Models are backtested before release, and each forecast sits next to your plan, so the gap to target shows early.
Questions in plain language
A language model translates each question into a query over governed metrics — never raw tables — and shows the query and definitions it used. If the governed data cannot answer a question, Lumen says so.
Recommended actions
For recurring decisions — reordering stock, moving budget, adjusting a price — Lumen proposes actions with an expected impact, shown as a range with its assumptions. A person approves, and the outcome is tracked against the estimate.
How it works
Built around five questions
What happened?
Governed metrics give one figure per KPI, with its definition, data freshness and quality status shown alongside.
Why did it happen?
Root-cause decomposition shows which segments drove the change and how much each one contributed.
What is happening now?
Continuous monitoring on fresh data flags anomalies as they form, not at the next reporting cycle.
What happens next?
Forecasts with uncertainty bands show where each metric is heading and how that compares with plan.
What should we do about it?
Recommended actions arrive with their expected impact and assumptions; people decide, and Lumen follows what happens afterwards.
Architecture
VARDA Lumen · Architecture
01 Sources
- ERP & CRM
- Warehouse & data lake
- Event streams
- Budgets & plans
02 Processing
- Orbit: ingestion & quality
03 VARDA Lumen
- Semantic metrics layer
- Analysis & forecasting engine
04 Outputs
- Plain-language answers
- Alerts & root-cause briefs
- Recommended actions
- Embedded analytics & BI
Use cases
Use cases
An illustrative case: a mid-sized retailer’s finance team closes the month knowing not only that gross margin fell, but which regions, categories and price changes caused it — before the management meeting, not after.
- Budget and forecast variance, decomposed by driver
- Month-end commentary drafted from governed figures
- Cash and working-capital forecasts with uncertainty bands
An illustrative case: a distributor’s sales operations team sees a region drifting from plan in the second week of the month, together with the products and channels behind it and a proposed stock rebalancing.
- Early warning when a region or channel drifts from plan
- Demand forecasts by product family, with prediction intervals
- Rebalancing and pricing actions with their expected impact
An illustrative case: a logistics operator tracks on-time delivery and cost per shipment from event streams, and the right owner is alerted as soon as a hub, carrier or shift moves outside its normal range.
- Near-real-time KPIs from operational event streams
- Root cause across sites, suppliers and shifts
- Alerts routed to the accountable owner, with context attached
Deployment options
Deployment options
On-premises
Lumen runs in your data centre beside the warehouse, with an open-weight language model on your own servers. No figure or question leaves your network — suited to banks and public institutions with strict data-residency rules.
Private cloud
Deployed on Kubernetes in your own cloud tenancy or a Turkish cloud region, reading from your cloud warehouse over private endpoints. Capacity scales with query and monitoring load.
Hybrid
Metrics and data stay on-premises, while alert delivery and embedded dashboards for partners run in the cloud. External language models can be allowed by policy for non-sensitive metrics only.
Technical specification
VARDA Lumen
- Metric definitions
- Versioned as code, with review and owner approval
- Query execution
- Pushed down to your existing warehouse as SQL; results cached by policy
- Anomaly detection
- Seasonal decomposition with robust thresholds, per metric and segment
- Root-cause analysis
- Contribution analysis across dimensions; price, volume and mix effects separated
- Forecasting
- Backtested statistical and ML models; 80% and 95% prediction intervals
- Natural-language interface
- Questions compiled to governed metric queries only; query shown with every answer
- Identity and access
- SSO via SAML 2.0 or OpenID Connect; row- and column-level access policies
- Interfaces
- Web app, REST API, embeddable components, email and Microsoft Teams digests
Integrations
- PostgreSQL
- Microsoft SQL Server
- Oracle Database
- ClickHouse
- Trino
- Snowflake
- Apache Kafka
- dbt
- Power BI
- Tableau
- Apache Superset
- Microsoft Teams
- Slack
Compliance & governance
- Designed to support KVKK obligations: personal data can be masked or aggregated before it reaches the metrics layer, and every access is logged.
- Designed to support the BDDK information-systems regulation for banks, with fully on-premises deployment, language model included.
- Designed to support ISO/IEC 27001-style controls: role-based access, single sign-on and a complete log of questions, queries and metric changes.
- Designed to support the GDPR principle of data minimisation where metrics cover EU residents.
Frequently asked questions
Frequently asked questions
01Our finance and sales teams report different revenue figures. How does Lumen resolve that?
Each KPI is defined once in the semantic layer — formula, filters, currency and calendar rules — and every dashboard, alert and answer reads from that definition. Changes are reviewed, approved and versioned, so any figure can be traced to the definition that produced it. Where two teams genuinely need different views, they become two named metrics rather than one ambiguous number.
02Can the language model invent numbers?
The model never calculates figures. It translates a question into a query against governed metrics; the query runs on your data platform, and the answer shows the metric, filters and query used. If a question cannot be mapped to a governed metric, Lumen says so instead of guessing.
03Do we have to replace our BI tools or our warehouse?
No. Lumen pushes queries down to the warehouse you already run and exposes governed metrics to your existing BI tools over SQL and a REST API. Dashboards can stay where they are; what changes is that they all share one set of definitions.
04Where does our data go when someone asks a question in plain language?
In an on-premises deployment, nowhere outside your network: the language model is an open-weight model running on your own servers. External API models can be enabled by policy, in which case only the question and metric metadata are sent — never raw records — and every request and response is logged.
05How is the expected impact of a recommended action calculated, and can we rely on it?
It is estimated from your own history — the results of similar past actions and the relevant forecast — and always shown as a range with its assumptions. It is an estimate, not a promise: a person decides, and Lumen records the actual outcome against the estimate so the models improve over time.
See VARDA Lumen with your own data.
Decision intelligence on governed metrics: one definition per KPI, anomaly and root-cause analysis, forecasts and recommended actions with expected impact.