Connectors
Your databases, files, APIs and Hadoop estate, connected in minutes.
Sovereign data & AI platform
From raw data to AI agents in production.On your infrastructure.
graal brings connectors, pipelines, lakehouse, notebooks, models and generative AI together in a single platform, installed on your infrastructure. Your teams and your AI agents run it with the same permissions, and every action is traced.
- On-premises
- Private cloud
- Qualified hosting provider
Demonstration
A node fails. Watch the pipeline recover.
conso-regionale-daily
- The pipeline runs
- A node fails
- The agent diagnoses and reruns
Job code
import pandas as pd # Aggregate consumption by region and by daydef agregation(df: pd.DataFrame) -> pd.DataFrame: return ( df.groupby(["region", "jour"]) .agg(conso=("conso_mw""conso_mwh", "sum")) .reset_index() )-- Aggregate consumption by region and by daySELECT region, jour, SUM(conso_mwconso_mwh) AS consoFROM curated.conso_regionaleGROUP BY region, jour;from pyspark.sql import functions as F # Aggregate consumption by region and by dayagregats = ( df.groupBy("region", "jour") .agg( F.sum("conso_mw""conso_mwh").alias("conso") ))Console
- # conso-regionale-daily · run-1285
- INFO ingest: 12 regions, 1,216,488 rows
- INFO cleaning: 0 duplicates
- ERROR agregation: KeyError: 'conso_mw'
- WARN retry 2/3: same error
- ◆ agent: the source now names the column conso_mwh
- INFO run-1286: agregation finished in 38 s
Event timeline
conso-regionale-daily · run-1285 → run-1286
- run-1285 started
- ingest
- cleaning
- agregation
- agregation · KeyError
- retry 2/3
- run failed · alert
- get_run_logs
- diagnosis: column renamed
- upload_code · fix
- approved by a human
- run_job · run-1286
- run-1286 succeeded
Animated illustration of a demonstration scenario, not a console screenshot: job, columns and identifiers are fictional.
The jobs of a project, their status and their last run. Rows marked “Agent” were created by an AI agent through MCP.
Real graal console interface, not retouched; fictional demonstration data (tenant energie-demo — people, projects and tokens are invented).
Agentagent · claudeLive
- create_jobprevisions/prevision-conso✓ created
- run_jobprevision-consorunning
- schedule_job0 6 * * *✓ scheduled
Service account agent-previsions · project previsions
Built on open standards
- Kubernetes
- Helm
- Apache Iceberg
- Trino
- Apache Spark
- Jupyter
- MLflow
- Keycloak
- PostgreSQL
- S3
- MCP
Platform
The whole data & AI lifecycle, on one platform
From the first connected source to the model served in production, without stitching ten tools together.
Low-code pipelines
Draw a pipeline: graal writes the Pandas or PySpark code, and it is yours.
Orchestration and compute
Jobs, workflows, cron and event triggers; distributed Spark and GPUs on demand.
Lakehouse and SQL
Iceberg tables on your S3 storage, federated SQL with Trino, catalog and lineage.
Notebooks
Jupyter and VS Code in the browser, with a coding assistant wired to your own model.
ML and generative AI
Experiments, model registry and serving; an LLM gateway and RAG over your documents.
AI agents
An MCP server so your agents create, run and schedule work, under your rules.
Governance
Per-project permissions, an audit trail of human and agent actions, encrypted secrets, costs and quotas.
Openness
Exportable code, open formats, REST API, MLflow and MCP: nothing locks you in.
How it works
From raw data to a model in production
One source, one secret, and it's wired.
- Databases, files, S3, SFTP and REST APIs
- Encrypted credentials, never written into code
- Schema detection and data preview
You draw, graal writes the code.
- Drag-and-drop visual editor
- Pandas or PySpark export that runs outside graal
- Quality checks at every step
The low-code editor canvas: every block carries its parameters, and the graph becomes code.
Real graal console interface, not retouched; fictional demonstration data (tenant energie-demo — people, projects and tokens are invented).
Schedule without writing a line of YAML.
- Cron, events and backfills
- Retries and alerts
- Live logs, run by run
The events of a workflow run, step by step.
Real graal console interface, not retouched; fictional demonstration data (tenant energie-demo — people, projects and tokens are invented).
Notebooks and SQL on the same tables.
- Jupyter and VS Code, single sign-on
- Trino on your Iceberg tables
- One catalog shared by every team
Every training run is tracked.
- Experiments, parameters and metrics
- Versioned model registry
- Works with the MLflow client
From model to endpoint, in one click.
- REST serving behind your SSO
- Scheduled drift monitoring
- An LLM gateway for your applications
Who did what, and what it costs.
- Permissions per project and per resource
- Audit of humans and agents
- Costs and quotas per project
Roles: what everyone can do, and on what.
Real graal console interface, not retouched; fictional demonstration data (tenant energie-demo — people, projects and tokens are invented).
AI agents
Your AI agents run the platform. Under your rules.
graal exposes an MCP server. Claude, or any compatible client, creates, runs and schedules jobs through a service account scoped to one project, and every action shows up live in the console.
- One service account per project
- Human approval for sensitive actions
- No destructive tool enabled by default
- Every call visible live
- Instant revocation
- Audit trail of humans and agents
You choose the model: hosted on your premises, or with the provider of your choice.
Agentagent · claudeLive
- list_runsprevisions✓ 12 runs
- get_run_logsrun-1284✓ read
- create_jobprevisions/prevision-conso✓ created
- run_jobprevision-consorunning
- schedule_job0 6 * * *✓ scheduled
Service account agent-previsions · project previsions
The “Agents” panel: who is connected, with which permissions, and every tool call live.
Real graal console interface, not retouched; fictional demonstration data (tenant energie-demo — people, projects and tokens are invented).
Deployment
Installed on your infrastructure. In a few commands.
One Helm chart, no cluster-admin rights and no custom resources: graal runs where your data lives.
- Helm
- OpenShift
- High availability
- Offline (air-gapped)
- GitOps
- Backup and restore
Your teams · SSO
◆ Your AI agents · MCP
graal · control plane
- Console
- REST API and MCP
- Identity · Keycloak
- Metadata · PostgreSQL
graal · execution
- Jobs and workflows
- Notebooks
- Distributed Spark
- Model serving
Your resources
- S3 storage
- LDAP / AD directory
- Image registry
- Observability
Your Kubernetes cluster · on-premises, private cloud or qualified hosting
On-premises
Your datacenter, your Kubernetes or OpenShift.
Private cloud
Your cloud, your account, your network rules.
Qualified hosting
Deployable with a SecNumCloud-qualified or HDS-certified hosting provider.
Use cases
What you build with graal
Energy
Forecast energy demand
Temperatures and regional consumption in a thermosensitivity model, refreshed every morning.
Industry
Predictive maintenance
Anticipate failures from sensor data, and alert before downtime.
Public sector
Sovereign document assistant
Query your internal documents with a model hosted on your premises.
Insurance
Fraud and claims
Flag unusual cases and explain every score.
Research
From datalab to production
Turn a notebook into a scheduled, governed and monitored job.
Cross-industry
Operations run by agents
Agents that watch runs, diagnose and restart them, under control.
Cross-industry
Take over an existing platform
Move from Saagie, Hadoop or a SaaS by bringing your Python and Spark jobs along.
Who it's for
One platform, five requirements
One platform instead of seven building blocks, on your infrastructure, at a predictable cost.
See pricing- Per-vCPU license, unlimited users
- Exported code and open formats
- Installed and upgraded with Helm and GitOps
Your data scientists explore, your data engineers industrialize, in the same tool.
See the platform- Notebooks, pipelines and orchestration together
- One shared catalog
- Every model tracked into production
No rights on the cluster, restricted pods, per-project permissions — agents included.
See security- Restricted pods (PSA restricted)
- Service accounts and human approval
- Every action audited
Data and processing stay within your infrastructure.
See sovereignty- No telemetry, no third-party service
- The model of your choice
- Personal data masking
One Helm chart, PostgreSQL, Keycloak and S3. Nothing else.
See the architecture- No custom resources
- OpenShift and high availability
- Offline installation
Trust
Trust you can verify
- Restricted pods (PSA restricted)
- SSO with OIDC, SAML and LDAP directories
- Permissions per project and per resource
- Secrets encrypted at rest (AES-GCM)
- Audit of human and agent actions
- No telemetry, no third-party service
Sovereign, strictly defined
- 01The platform and the data run on the infrastructure you choose.
- 02No third-party service is required for it to work: no SaaS, no telemetry, no external fonts or captcha.
- 03The vendor, which also integrates and supports the platform, is a French company (GRAAL.SYSTEMS SAS).
- 04Pipeline code is exportable and formats are open (Parquet, Iceberg, SQL, MLflow API).
Get started
Start with an 8-week pilot
One use case end to end, on your data, in your infrastructure. Then you decide, on evidence.
Week 1
Scoping: one use case, clear success criteria
Weeks 2–3
Installation on your infrastructure
Weeks 4–7
The use case end to end, on your data
Week 8
Review and decision
Frequently asked questions
Do we need a Kubernetes cluster?
Yes. graal installs on your Kubernetes or OpenShift, on-premises or with a hosting provider. If the cluster still has to be built, we help you.
Which model drives the agents?
The one you choose: a model hosted on your premises and served by graal's LLM gateway, or the provider of your choice. graal speaks MCP, an open protocol.
Is graal SecNumCloud-qualified or HDS-certified?
Those labels qualify hosting offers, not software. graal deploys with a SecNumCloud-qualified or HDS-certified hosting provider, just as it does in your own datacenter.
Does our data leave our infrastructure?
No. The platform and the data stay with you. If you connect an external model, it only receives what the tools it calls return.
Can it be installed without Internet access?
Yes. graal installs offline, from your own image registry.
How is pricing calculated?
An annual license per vCPU allocated to workloads, unlimited users, no usage-based billing.
Who provides support?
The vendor. graal Services cover integration, migration, training and operations.
What if we want to leave?
Pipeline code is exportable and formats are open: you leave with your work.
Let's see graal on your data
Data teams
Watch the demo (14 minutes)CIOs and CISOs
Get the architecture packPublic buyers
How to buy graal