Research project on safe agentic tool use
This project aims to run the core router agent locally, so sensitive data stays in your local environment.
The goal of this research is continuous learning with a trainer agent that updates the memory database and fine-tunes the neural network every day. No need for flowcharts or detailed task descriptions by a user.
This system is already capable of browsing, comparing, and updating text and Excel files, as well as analyzing images and screenshots. We are currently testing and training more complex use cases such as CAD modeling.
Most day-to-day tasks don't require a powerful frontier model. By running the core router agent locally, all company data stays securely within your own environment. Nothing leaves your system unless you explicitly allow it. For more advanced tasks, the router can still leverage a frontier model — but only with carefully filtered or redacted information — ensuring sensitive data always remains protected.
By combining AI with a structured SQL database, we get the best of both worlds: intelligent reasoning with full control. The ROSA system learns workflows through targeted LoRA fine-tuning, while critical business data remains securely stored and managed in the database. This prevents errors such as mixing up customer names, part numbers, or internal data. Instead of guessing, the AI retrieves real-time information directly from trusted sources like ERP systems or email. This approach also makes it faster and more cost-efficient to teach the system new tasks.
Every action the agent can perform is governed by clear authorization levels. Low-risk actions, such as browsing or gathering information, can run freely. Medium-level actions, like moving files, are only executed when they match predefined rules and approved structures. High-impact actions, such as deleting files, always require direct user approval. By storing approved actions in the database, you define exactly how autonomous the system can be — giving you innovation without sacrificing control.