Enclave
Launch, serve and fine-tune any model inside your own cloud.
Managed model serving and fine-tuning
Launch and serve any open-source or custom model on your own GPUs, from chat and embedding models to your own fine-tunes. Fine-tune and evaluate models on the compute you choose, use pgvector for retrieval, and reach company data through LakeGate. Eligible endpoints scale down when idle, and sensitive work stays in your region.
- Works alongside
- Public model APIs, for everyday work that isn’t sensitive.
- Can replace
- Private hosted endpoints for open models, and hosted vector databases.
- Any engine to any engine
- An agent pulls context from pgvector and joins it with ClickHouse facts in one governed session.
AgentBay
A safe place for your agents to work.
Managed agent sandboxes and MCP
Isolated sandboxes for agent code and tools, with data access through LakeGate, where policy decides each table and column. Agents built with any toolkit can query any engine, act and report back through Hyperlake MCP.
- Works alongside
- The agent toolkits and MCP clients you already use.
- Can replace
- Hosted code sandboxes, and database credentials handed to agents.
- Any engine to any engine
- Sandboxed agent code runs a query. LakeGate routes it to Postgres or ClickHouse, and policy filters the rows.
Analyst
An AI analyst that answers from governed company data.
AI analyst on your own data
Velocity Lakehouse, Enclave and AgentBay together. The semantic layer, grants and policy decide what the analyst may see, private models answer, sandboxes run the analysis code, and monitors watch data quality.
- Works alongside
- Your BI tools and existing warehouses, through LakeGate and catalogs.
- Can replace
- Chat-with-your-data tools that copy data out to a vendor, and text-to-SQL add-ons sold per seat.
- Any engine to any engine
- A question resolves through the semantic layer, LakeGate routes it to ClickHouse or Trino, and a private model writes the answer.
Predict
Serve the models you already train, next to the data.
Managed ML model serving
Feature models in dbt, model serving for sklearn, XGBoost, LightGBM, ONNX and Triton models, prediction logs in ClickHouse, and data-quality monitors.
- Works alongside
- Training wherever you do it today.
- Can replace
- Managed real-time endpoints for classic ML models.
- Any engine to any engine
- dbt builds features in Iceberg, the model scores them, and predictions are written back to ClickHouse and Postgres.
Physical AI Forge
Robotics and autonomy pipelines on your own GPUs.
Managed NVIDIA OSMO on your GPUs
NVIDIA OSMO orchestrates the physical-AI loop on your GPU pools: synthetic data, training, reinforcement learning and software-in-the-loop testing. Telemetry lands in ClickHouse, curated datasets live in Iceberg, and trained models are served through KServe.
- Works alongside
- The simulation and robotics tools you already use, and existing simulation farms.
- Can replace
- Hand-built GPU job scripts and queue systems for robotics teams.
- Any engine to any engine
- Fleet telemetry in ClickHouse is filtered into Iceberg training sets, OSMO trains and tests, and the model is served to the fleet.