VedaAnalytics
Sovereign Self-Hosted Analytics Platform with Local AI
A self-hosted analytics workspace: connect 25 kinds of source, stream into an Iceberg lakehouse, explore with SQL or plain-language questions answered by models on your own GPUs, and publish dashboards with forecasts.
- Ask questions in plain words: AI writes read-only SQL and answers with a chart
- 25 kinds of source, change data capture and drag-and-drop flows into an Iceberg lakehouse
- No-code dashboards with drill-down, forecasts and unusual-point detection
- Runs on your own servers with local models: your data never leaves
Overview
VedaAnalytics puts the work of a data team in one self-hosted workspace that people without SQL can use. Connect sources, load or stream data into an Apache Iceberg lakehouse, profile and query it, build dashboards with filters and drill-down, forecast trends, flag unusual points and share the results.
People ask questions in plain words. A model running on your own GPUs writes the SQL, every query is checked to be read-only before it runs, and the answer comes back as a chart. Administrators choose which models serve each part of the platform in the Model Hub, so no data leaves your servers for an external AI service.
Key capabilities
-
Ask Veda
A plain-language question becomes SQL from a local model, is checked by the SQL guard, shown before it runs, then approved, run read-only and charted automatically with findings.
-
VedaAI Copilot
Natural language to SQL grounded on the live schema and sample rows of the datasets you choose. Every answer is validated as a single read-only SELECT, with automatic repair rounds and an audit trail of every step.
-
Iceberg lakehouse
Apache Iceberg tables in S3-compatible object storage, an Apache Polaris REST catalog and Trino queries, with typed columns and a snapshot per write. Read-only for users in three independent layers.
-
Connections
25 kinds of source including PostgreSQL, MySQL, SQL Server, Oracle, SAP HANA, IBM Db2, Snowflake, BigQuery, Redshift, ClickHouse, MongoDB, Cassandra, Elasticsearch, S3, Azure Blob, Google Cloud Storage, SFTP, HDFS, REST APIs, Google Sheets and Kafka. Secrets are encrypted and never returned.
-
Change data capture
Continuous replication of PostgreSQL, MySQL, MariaDB, SQL Server, Oracle and MongoDB tables into the lakehouse: an initial snapshot, then every insert, update and delete within seconds, in mirror or history mode.
-
Flow designer
Drag-and-drop data flows from sources through transforms (select, filter, join, group, deduplicate, union, SQL and Python) to lakehouse tables, run on the built-in engine, Apache Spark or Apache Beam with the CPU, memory and GPUs you assign.
-
Pipelines
Versioned pipelines from uploads, storage folders, database tables or Kafka topics, with field mapping, quality checks with quarantine, test runs on samples, schedules and run history.
-
Explore and query
A dataset explorer with schema, preview, profiles and snapshot history; a SQL editor with completion and an AI bar to write, fix, explain and improve queries; dataset diagrams that find the joins between tables.
-
Dashboards
A no-code chart builder with 11 chart types, drag-and-resize dashboards, filter controls, click-to-filter and time drill-down, Holt-Winters forecasts with 95% bands, unusual-point detection and plain-language findings.
-
Share and export
Read-only share links and embeds; export every row to CSV, Excel, Parquet, JSON Lines or TSV.
-
Model Hub
GPU inventory and VRAM per model, the whole Ollama model library, replicas placed on chosen GPUs, routing of each platform task to a model, least-busy load balancing, and models pinned in memory across restarts.
-
Backups
Daily encrypted backups of the control plane, catalog, warehouse, secrets and lakehouse bucket, with a weekly automated restore test.
-
Built-in help
A detailed, searchable, offline help page for every screen, with diagrams, option tables and troubleshooting.
Architecture
VedaAnalytics is a Go API server and a React web interface, run as a systemd service. All analytical data lives in object storage as Apache Iceberg tables registered in Apache Polaris and queried by Trino; the API is the only writer, and every user, dashboard and AI query reaches Trino only through the SQL guard. Local models run in Ollama containers, each pinned to a GPU and sharing one model store.
-
Workspace
- React web interface
- Explore, Ask Veda, dashboards, flow designer
- Command palette and built-in help
-
API and protection
- Go API server
- Application firewall and rate limiting
- SQL guard: read-only, single statement
-
AI
- VedaAI Copilot
- Model Hub: runtimes, replicas, task routes
- Ollama runtimes pinned to GPUs
-
Ingestion
- Connections (25 kinds of source)
- Pipelines and flows: built-in engine, Spark, Beam
- Change data capture with Debezium and Kafka
-
Lakehouse
- Apache Iceberg tables
- Apache Polaris REST catalog
- Trino (read-only for users)
- S3-compatible object storage
-
Control plane
- PostgreSQL 16 (bundled or external)
- Audit trail
- Encrypted secrets and backups
Technical specifications
| Packaging | Go API server and React interface as a systemd service; lakehouse services and model runtimes in Docker containers; one-command installer (setup.sh) and full uninstaller |
|---|---|
| Operating system | Ubuntu or Debian |
| Table format and engines | Apache Iceberg; Apache Polaris catalog; Trino; bundled SeaweedFS object storage or your own S3, GCS, catalog or Trino |
|---|
| Engines | Built-in engine (DuckDB), Apache Spark and Apache Beam, each in its own container with assigned CPU, memory and GPUs |
|---|
| Broker | Bundled Kafka broker; topics stream into lakehouse tables |
|---|
| Models | Any model from the Ollama library, managed in Model Hub; ships with a coding model for SQL and an embedding model |
|---|---|
| Hardware | NVIDIA GPUs (detected and configured by setup), or CPU-only with slower AI answers |
| Transport | TLS 1.3 on every connection, including between internal services |
|---|
| Verified | Google Cloud G2 (1× NVIDIA L4 24 GB, Ubuntu): fresh install in under four minutes |
|---|
Deployment
-
On-premises
Your own Ubuntu or Debian server, with or without NVIDIA GPUs.
-
Cloud (your account)
A GPU instance in your own cloud account.
-
Single node or your own services
Everything bundled on one server, or connect your own S3/GCS storage, Iceberg catalog, Trino cluster and PostgreSQL.
Requirements
| Server | Ubuntu or Debian; an NVIDIA GPU is recommended for AI features |
|---|---|
| Software | Installed by setup.sh when missing: Docker and Compose, NVIDIA Container Toolkit (GPU servers), Go, Node.js |
| Network | HTTPS on port 443; all internal services bound to loopback |
Supported integrations
-
Relational databases
Sources
PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, SAP HANA, IBM Db2.
-
Cloud warehouses
Sources
Snowflake, BigQuery, Redshift, ClickHouse.
-
NoSQL and search
Sources
MongoDB, Cassandra, Elasticsearch / OpenSearch.
-
Object and file storage
Sources
S3-compatible, Azure Blob / ADLS, Google Cloud Storage, SFTP, FTP/FTPS, HDFS, NFS/SMB folders. Files as CSV, TSV, JSON, Parquet or Excel.
-
APIs and streams
Sources
REST/JSON APIs, Google Sheets, Kafka topics, HTTP ingest endpoints.
-
Identity providers
Identity
OpenID Connect and SAML 2.0 single sign-on (Keycloak, Entra ID, Okta, Google, Ping, ADFS), LDAP and Active Directory, SCIM 2.0 provisioning.
Security & governance
-
Read-only analytics
User, dashboard and AI-generated SQL passes a guard that allows a single read-only statement; Trino rules and least-privilege catalog and storage credentials enforce the same independently.
-
Human approval for AI
AI-written queries are shown before they run and execute only after a person approves them; every step is audited.
-
Sign-in
Two-factor sign-in with authenticator apps and recovery codes, single sign-on, LDAP, lockout after failed attempts, session revocation and a forced password change at first sign-in.
-
Encryption
TLS 1.3 on every hop with an internal certificate authority; connection secrets encrypted and never returned; encrypted backups.
-
Local AI
Models run on your own servers; questions and data are not sent to external AI services.
Use cases
-
Business teams
Self-service analytics
Ask questions in plain words and get charts without waiting for a data team.
-
Government & public sector
Sovereign data platform
A complete lakehouse, BI and AI stack that runs entirely inside your own infrastructure.
-
Enterprise IT
Operational replication
Replicate production databases into the lakehouse within seconds for reporting without loading the source systems.
Documentation & developer resources
Release history
- VedaAnalytics: A Unified Self-Hosted Platform for Data, AI and Analytics VedaAnalytics: A Unified Self-Hosted Platform for Data, AI and Analytics
Frequently asked questions
Does VedaAnalytics send our data to an external AI service?
No. Models run on your own servers through the Model Hub, and AI-written queries run read-only after a person approves them.
Do we need a GPU?
A GPU is recommended for fast AI answers. VedaAnalytics also runs on CPU-only servers; setup detects this and warns that AI answers will take longer.
Can we use our existing storage and query engine?
Yes. Use the bundled single-node services, or point VedaAnalytics at your own S3-compatible or GCS storage, Iceberg REST catalog, Trino cluster and PostgreSQL.
Related technical articles
Article · VedaAnalytics
VedaAnalytics – Engineering Notes
VEDACLOUD.AI / PRODUCT ENGINEERING A four-part technical blog series for the VedaAnalytics launch SERIES CONTENTS Inside VedaAnalytics: A Self-Hosted Lakehouse with Separate Control, Data and Model Planes An…