Radix
Radix — a 36-month AI research and development program no one has to risk alone
Partnership in progressProject supported by Finep, Brazil's federal innovation agency
The problem
In engineering companies, project management is critical. Delays, billing deviations and decisions made without the right knowledge are expensive, and that knowledge is usually scattered across people, documents and systems.
Radix decided to tackle this with a 36-month, R$ 28.4 million research and development program supported by Finep: a project management copilot based on a generative AI multi-agent system, developed in Brazil and built for Portuguese. In the words of João Carlos Chachamovitz, CEO of Radix: "We are developing a technology that combines generative AI and multi-agent systems to support decisions in highly complex environments." (translated from Portuguese)
Radix has its own AI team and could have run the project alone. But a long program on a frontier topic concentrates a lot of risk in a single team. Radix sought technical partners to reduce the uncertainty, and built a partner network with Bolder, InnoVox, COPPE/UFRJ, PMI and 42 Rio, a format Finep itself values.
InnoVox's role
InnoVox is part of the project's execution network as a partner for applied research and engineering of the AI agents. The system was designed from the start to grow to more than 20 specialized agents.
Discovery & Diagnosis
Before building any agent, we carried out technical feasibility research on what had been specified. That research changed the architecture before construction: we identified a shared knowledge base that every agent can consult to make better decisions, instead of each agent reinventing its own. We also defined where generative AI should act and where the calculation has to be deterministic and auditable.
Proof of Concept — in progress
Every major decision was made on evidence, not on tool preference:
- a structured comparison of architecture alternatives, with decisions approved by technical leadership;
- systematic literature reviews, feeding into the scientific publication planned in the program;
- tests with real project documents before implementation.
Engineering & Production — upcoming cycles
Agent implementation is moving forward cycle by cycle, with acceptance testing and indicators set in the contract for the final phases of the program.
What this proves
Applied AI & Knowledge Engineering: applying AI to specialized knowledge with research rigor, knowing where generative AI adds value and where it adds risk.
And our method inside a formal research and development program: start with the problem, reduce uncertainty with evidence, and only then build.