I’m very happy to share one of the most meaningful pieces of work I’ve done in my career so far.
For the last few months, we have been working on a system for multi-agent orchestration and model routing with one clear focus:
Reducing energy consumption without compromising quality.
In the AI world, most of the attention around energy goes to model training, model development, and model serving.
But there is another layer that is often overlooked:
The energy consumption of the systems we build around LLMs.
Our model routing architecture at VDF AI works across 6 layers, can operate in 4 different modes, and continuously learns through LinUCB.
In our tests, we measured a 81–95% reduction in energy consumption, without degrading output quality or changing the expected system behavior.
Seeing that result was honestly one of the most rewarding moments of this journey.