01
Conversational composition
Each chat owns one layer and assembles the result by combining ops (fetch / add / filter / aggregate / visualize). Reference upstream commits as REF; branch any commit to try alternatives without losing the original.
01 / Product
A workspace that turns questions into governed Iceberg layers.
Lanikaia is a conversational AI platform for modern data teams. Domain experts, analysts, engineers, and scientists share the same project, build the same pipeline, and read the same artifacts in the language each role prefers.
02 / How it works
The unit of work is a chat that builds a layer.
Open a chat against a layer. Ask in plain language. The agent generates Python that runs on Polaris-managed Iceberg, writes the result, and explains what it did.
Each chat keeps a working draft and a list of commits. Branch a commit to try a different aggregation. Promote a commit to a workflow when it is the version you want to ship.
- 01ingestLocal files, HTTP, dbt SL, SAP, BO
- 02foundationCleaned, typed, schema-stable
- 03enrichedJoined, derived, dimensional
- 04aggregateGroup-bys, daily / monthly rollups
- 05cumCumulative measures across time
- 06joinCross-domain composition
- 07rollingWindowed and lagged features
- 08metricsdbt Semantic Layer / OSI surface
- 09analyzeStatistical and exploratory analysis
- 10testingHypothesis tests, validation
- 11predictML / classification, regression, time series
- 12optimizeSearch, scenario, what-if
- 13visualizePlotly + ECharts charts and dashboards
- 14workflowPromotion, schedule, monitor
03 / Who uses it
Four lenses on the same project.
Domain expert
Asks the question in their own language. Reads the explanation, the chart, and the underlying number side by side. Decides without learning SQL.
Analyst / BI
Composes intermediate tables in chat, branches alternatives, and pins what works. Each commit is a reproducible artifact, not a one-off.
Data engineer / SRE
Promotes a chat to a workflow. Inspects layers, lineage, and code. Diagnoses a broken pipeline by layer in five minutes, not five hours.
Researcher / scientist
Iterates models inside the same chat that owns the input table. Uses Polars, scikit-learn, LightGBM, statsmodels, UMAP. Versions every run.
04 / What it does
Six capabilities, no decoration.
02
Real Python under the chat
The agent generates Python that runs against Polaris-managed Iceberg tables. Output is auditable code and a versioned table, not a screenshot.
03
One project, four lenses
Domain experts read prose. Analysts read tables. Engineers read code and lineage. Scientists read notebooks. Same project, four lenses on the same artifacts.
04
Publish as API or MCP
Promote a dataset or function to a versioned HTTP API or an MCP server. Other agents and applications consume the result without going through the UI.
05
Governed by default
Polaris catalog scopes every project. Identity and permissions flow from your IdP. Reads, writes, and grants are audit-logged.
06
Lossless intermediate
Every intermediate table is an Iceberg snapshot, not a CSV. Schema changes do not silently break downstream. Time-travel and lineage are first-class.
05 / Next