Most AI models are evaluated as if they will run in ideal conditions. But real edge environments don’t work that way.
That tension sits at the core of E*3, the project developed by Jorge Martínez during the Second Open Call of dAIEDGE.
Instead of asking a single question, “which model is more accurate?”, E*3 reframes the problem entirely: what if accuracy is only one part of the equation?
The project introduces a multi-objective perspective for text embeddings, where latency and energy consumption are treated as first-class constraints alongside performance. This is especially relevant at the Edge, where hardware limitations define what is actually feasible.
Under this framework, embedding models are no longer compared only by benchmark scores, but by how they behave under real deployment conditions. The result is a more practical way to decide which models fit which edge scenarios, and why.
During the programe, Jorge Martínez worked in close collaboration with Synopsys, whose expertise helped shape both the technical direction and evaluation methodology of the project. We would like to sincerely thank Synopsys for their role as hosting institution.
One of the key enablers for this work was the dAIEDGE-VLab environment, where different models and configurations were tested under realistic hardware constraints. Rather than validating ideas in isolation, the project was continuously stress-tested against conditions closer to real deployment, which helped refine both assumptions and outcomes.
Beyond the technical development, the experience also highlighted something equally important: access to shared infrastructure, expert mentoring, and a European ecosystem that makes it possible to move from theory to usable systems faster.
E*3 is a reminder that efficiency is not a trade-off—it is a design dimension.