Talk to your data
Make every question decision-ready.
One governed platform for analytics, dashboards, and deterministic data intelligence.
Three pillars. One governed system.
Studio is the intelligence surface. The Data Language Model resolves supported questions deterministically. The Rust Engine provides the vectorized analytical foundation beneath both.
Kaveon Studio
Questions, SQL Lab, governed dashboards, chart building, and operational exploration.
Data Language Model
Compiled dataset semantics and deterministic NL→SQL for supported question classes.
Kaveon Engine
Distributed Arrow execution over Parquet and Delta tables on local storage, with exact memory accounting. Qualified on a cluster; broader cloud and format qualification remains staged.
The work, in one system
Deterministic answers, distributed analytics, transactional product state, and governed ways to share the result.
Ask your data
Type questions in plain English. The Data Language Model resolves supported questions against compiled dataset context, generates auditable SQL, and renders the answer as an interactive chart — with no hosted model call.
See the signal
Turn governed answers into interactive charts, maps, and decision-ready views.
Query without friction
SQL Lab with autocomplete, multi-tab workspaces, history, and result caching.
Dashboards That Tell Stories
Drag-and-drop canvas with cross-chart filtering, shared filter bar, auto-refresh, and one-click publishing.
Connect your sources
Register lake catalogs and supported SQL sources in one governed workspace; each query uses its selected execution path.
Governed semantics
Define dimensions, metrics, and joins once, then reuse them across questions, charts, and dashboards.
Transactional product state
KaveonDB keeps datasets, charts, dashboards, saved queries, DLM artifacts, and audit records behind one transactional boundary.
Governed access
Microsoft Entra ID, role-aware access, and customer-controlled data boundaries.
Own your deployment
Run locally with Docker or deploy the Studio and Engine as separately qualified private-cloud services.
Ask anything. Get answers.
Top country by Carbon Intensity (10 results)
*...and 5 more*
From source to decision.
Connect once, define meaning once, then explore through natural language, SQL, and reusable visual intelligence.
Connect
Register governed sources without moving the underlying data.
Model
Define reusable dimensions, metrics, relationships, and business meaning.
Ask or query
Use deterministic DLM resolution or write SQL directly in Studio.
Execute
Run vectorized analytical work through the Engine or a selected source.
Visualize
Turn answers into charts, dashboards, and operational decisions.
Why deterministic, not generative?
Every NL→SQL product today sends your schema to an LLM. Kaveon doesn't.
Bounded Semantics
Supported questions map to compiled dataset metadata and schema boundaries instead of unconstrained token prediction.
Predictable Runtime
Compiled rule-based parsing avoids a token-generation pipeline. Published latency claims require a reproducible benchmark context.
Explicit AI Boundary
The deterministic DLM path makes no model call. Optional hosted-AI features remain separately configured and governed.
The Data Language Model
A self-compiling semantic layer that turns your schema into a deterministic question-answering engine — no training, no fine-tuning, no LLM.
Schema in, answers out
Register a dataset and the DLM compiles itself — indexing every column value, mapping synonyms, and precomputing metric rollups across every dimension. Questions resolve to SQL through deterministic pattern matching, not token prediction.
Zero-scan change detection
Every precomputed answer carries a validity score in [0, 1]. The algorithm detects data drift from database catalog counters — never by re-querying your tables.
Exponential decay with a 6-hour base half-life. Unused context ages out gracefully; heavily-used context decays faster so your most-relied-upon answers stay freshest.
Reads pg_stat_user_tables.n_mod_since_analyze — a counter the database maintains for free. Detects data drift without scanning a single row.
Hot elements get a shorter effective half-life. The answers your team relies on most are kept the freshest — the algorithm learns from access patterns.
Adaptive Context Routing
Every question is routed through a per-element freshness scorer that reads source statistics and artifact metadata — not a full data rescan. Fresh context can answer directly from the DLM. Stale elements trigger targeted live queries and refresh the context for next time.
Build dashboards that tell stories
World maps, KPI cards, trend lines, donut charts — drag, drop, publish.
arena_elo, mmlu, input_cost
FROM ai_benchmarks.leaderboard
WHERE arena_elo IS NOT NULL
ORDER BY arena_elo DESC
LIMIT 10;
VS Code in your browser
Monaco editor with SQL autocomplete, syntax highlighting, multi-tab sessions, query history, and result caching across supported connectors.
Open SQL Lab →Benchmarks
All 43 ClickBench queries,
timed on Kaveon.
The 43 upstream ClickBench statements over 99,997,497 rows of the public hits table, on 3 Standard_D4s_v3 worker nodes with 3 GiB per query per worker. Each round starts cold and runs every statement 3 times; a statement’s figure is the median of its round medians over 3 rounds.
ClickBench, 43 statements on Kaveon: 9 under 1 second, 21 under 10 seconds; slowest q33 at 184 seconds.
- Rows scanned
- 99,997,497
- Rounds
- 3, cold, 3 executions each
- Cluster
- 3 × Standard_D4s_v3 workers, 3 GiB per query per worker
- Recorded
- 17 September 2026
- Engine build
- 8580e4ea3728
- Record
- Round records on GitHub