Agents built with you.
We design your agents with your teams, in a visual interface: data sources, tools, steps, access rights. The goal: a reliable agent whose every decision stays explainable, auditable, and under your control.
- Regulatory reporting
- Document analysis
- AI for public services
What you can do
Governed by your rules
Tools, models, behavior, access rights: every agent is governed by your rules. Zero shadow AI.
From documents to decisions
Powered by ODIN: complex documents (PDF, PPT, DOC, video, audio) are perceived, structured, and understood natively for AI.
Every AI decision, explained
5-layer monitoring covers infrastructure, model, agent, RAG and behavior. Every call is logged and every source cited.
Documentation
An agent built with you. Governed in production.
Devana.ai is used to build specialized agents in critical sectors, as at BIL and within the Luxembourg government. We design and test them with your teams, following the Devana Protocol©; you then monitor and govern them.
We set up each agent with you in a visual editor: the agent at the center, linked to its settings.
- AI model
- An assistant guides the choice in three steps: area of expertise, data processing requirement, model. Local models are part of the choice.
- Identity
- Two modes: creative, or limited to your knowledge base for accurate, controlled answers. The system identity, the temperature, and the reasoning level are adjustable.
- Restrictions
- Prompt restrictions validate the questions asked of the agent before it answers.
- Knowledge bases
- The agent is attached to one or more bases, your own or those a team shares. A base brings together its documents, its sources, and the agents that use it.
- Connectors
- Google Drive, Dropbox, OneDrive, SharePoint, Jira, Gmail: data syncs from your tools. SharePoint accepts Kerberos service accounts.
- Web sources
- A URL is enough: the content of the site becomes available to the agent.
- Tools
- Your business tools, MCP servers, and other agents can be attached to the agent. A dedicated model, instructions, and temperature drive how they run.
- Business vocabulary
- A glossary defines your terms and acronyms; replacement words correct the answers automatically.
- Chat and audience
- Welcome message, suggested questions, display of sources, voice mode, and the audience the agent addresses.
Reliability is checked before launch: the agent goes through acceptance testing with you.
- Challenges
- A challenge is a set of test questions, run against one or more agents. Every run is kept.
- Test chat
- The agent can be tested from the editor, by text or by voice, before it goes into service.
- Duplicate, export, import
- An agent can be duplicated, exported as JSON, and imported into another environment.
- Production mode
- Production mode is switched on agent by agent.
In production, each agent has its own dashboard, which can be exported.
- Usage
- Conversations, users and messages per day, activity by hour, response time and average conversation time.
- Reliability
- Reliable conversations, evolution of response scores, error rate and message blocking rate.
- Sources
- The documents with the highest and the lowest RAG Score, and the most cited sources.
- Tools
- The distribution of tool calls, agent by agent.
- Consumption
- Tokens per day, per model and over six months, with a monthly cap and an alert per agent.
- Messages and logs
- Charts, messages and logs are in one place, and can be exported.
Administration sets the framework for the whole organization.
- Users and teams
- Manage users and teams; agents and knowledge bases are shared by team or by email.
- Authorized models
- Choose the models available, add custom models and moderation templates.
- Tools and connectors
- Custom tools, connectors and MCP servers are declared at organization level.
- Message retention
- The lifespan of messages is configurable: they are deleted automatically at the chosen term.
- API access
- API keys and OAuth clients to integrate your applications.
- Personal data
- Each user can download their data or delete their account.
In production in critical sectors
- BILBerry, the agent of Banque Internationale à Luxembourg: 300,000 users in production, 98% answer reliability, deployed within the bank’s infrastructure.See the BIL case
- CTIEDocument agents used every day across the ministries of Luxembourg, on a sovereign infrastructure operated by the CTIE teams.See the CTIE case
Runs on your premises.
Where this application runs, and the technologies that come with it.
- On-premises
- On your servers, with a configuration sized for you.
- DIWY
- Available on DIWY, depending on your configuration. Discover DIWY
- Requirements
- Devana.ai runs on Devana OS. Models, data sources, and approved tools are configured with you, on your servers or on DIWY.
Explore further
Technologies that work together.