Wakeline, based in Düsseldorf, has just closed a €2.1 million pre-seed round led by TechVision Fonds, with participation from neoteq ventures. Founded in 2025 by Tim Gülke, Jan Böggering, Simon Sprünker, and Merten Tiedemann, the team aims to tackle one of the obvious limitations of current AI systems: their reliance on periodic training with historical datasets.

Why intermittent updates aren’t enough

Most models today are trained on past data, then frozen and deployed. Updates happen at fixed intervals, when enough new data accumulates to retrain the entire system. While this approach has delivered impressive results, it creates a gap between what the model “knows” and the shifting reality it operates in. In sectors like manufacturing, logistics, or cybersecurity, conditions change fast, and waiting for a retraining cycle can mean basing decisions on outdated information.

Wakeline proposes a paradigm shift: merging learning and deployment into a continuous flow. The goal is to let systems incorporate new knowledge while they run, without shutting down or sending data to a centralized cluster. The architecture, inspired by biological learning mechanisms, is not tied to proprietary models or hyperscale cloud infrastructure — an aspect that could attract organizations focused on data residency and autonomous infrastructure management.

Cloud independence: a matter of control and costs

Being free from hyperscaler lock-in goes beyond a technical preference. For many companies, constantly moving sensitive data to the cloud for retraining brings compliance risks (GDPR, sector regulations) and non-trivial operational costs. A system that can learn incrementally on the same infrastructure that runs it — whether on-premise, at the edge, or in a hybrid setup — reduces dependence on external pipelines and simplifies data governance.

Of course, running continuous learning locally requires adequate computing power. This is not a “lightweight” technology: maintaining an updated knowledge state without performance loss demands performant hardware, possibly with GPU acceleration, and efficient VRAM management strategies. But if the cost of periodic cloud retraining is replaced by steady, controlled on-premise consumption, the TCO could prove favorable in the medium term, especially for highly variable workloads.

Looking ahead (and at the trade-offs)

With the new funding, Wakeline will invest in platform development, team expansion, and initial cross-industry commercial activities. The promise of AI that evolves in real time is alluring, but technical questions remain: how to prevent incremental learning from introducing drift or catastrophic forgetting? What validation mechanisms ensure the model doesn’t degrade with experience? And, crucially for those evaluating self-hosted deployment, what are the minimum hardware requirements?

AI-RADAR follows these topics closely because they touch on core issues of on-premise AI deployment: data sovereignty, model control, and update transparency. Wakeline’s approach, if proven in practice, could add another piece to an AI ecosystem independent of large cloud providers, without sacrificing adaptability. For now, it’s a strong signal that research is pushing beyond the “train-deploy-freeze” boundaries we’ve grown accustomed to.