AI RESEARCH LAB
Wakeline’s AI learns while it runs, with no pretraining, no retraining and at CPU cost. We first deployed it in energy markets. Now select companies across other industries are testing it.
THE PROBLEM
Today’s AI systems are powerful during training and static after deployment. The world keeps changing, and the system’s understanding of it decays.
That has three costs. The system responds slowly to new signals, because it cannot take them in until it is retrained. Operational risk rises as conditions drift away from the training data. And it takes more and more compute and complexity, in retraining runs, pipelines and monitoring, to hold performance steady.
OUR THESIS
Wakeline is an AI research lab in Düsseldorf and Hamburg, bringing together scientists from physics, nonlinear dynamics, computer science and neuroscience. We built an architecture that learns while it runs, with no pretraining, no retraining, and one instance per asset on an ordinary CPU.
Intelligence is a process in time: expect, act, observe, revise. A system must learn from the consequences of its actions and revise what it knows as conditions change. Our thesis is that this ability to keep learning during operation is foundational to general intelligence.
The same learning mechanism has been applied across three different environments. In the European power market, it learns from live signals to forecast prices and inform operational decisions. In a simulated stock market, it learned hidden action costs from trading outcomes, without a reward function. In a simulated road network, it learned to anticipate delays and improved how it acted on its forecasts.
These applications test connected capabilities: learning what happens, learning what actions yield, and using that experience to make better decisions while operating. Energy is our first production deployment; the simulations demonstrate capabilities we are now seeking to test more widely.
Market Edge How one architecture works across environmentsRESEARCH
Deployed
Market Edge has forecast day-ahead electricity prices continuously in production since April 2026. It learns while operating, currently ingesting 83 parallel input streams. Battery storage operators and energy traders use its forecasts to plan their operations.
Backtested
Historical backtest, Market Edge: 14 percent higher cumulative trading P&L than the TabPFN baseline across 2,097 days in the German-Luxembourg power market. The advantage concentrates in regime shifts. A backtest result, not live trading returns.
Demonstrated
In a simulated stock market, an instance acted and learned what its actions cost: over 2,177 trades it recovered a hidden 25 percent tax schedule from the consequences of its own trades, with no reward function, and recovered it again each time the rule changed.
Demonstrated
In a simulated road network, an instance learned to predict delays and to choose a route before the delay happened, and kept improving how it acted on its own forecasts as conditions changed. We are looking for design partners to take this into operational use.
OUR TECHNOLOGY
No model is trained in advance. Setup connects the architecture to operational data via API or webhooks, with someone who understands the values. A cold start needs a history of a few hundred data points.
Learning is self-supervised: each new observation grades the previous prediction. The instance learns directly from this feedback as operational data arrives, with no labelled data needed to guide its learning.
Nothing is trained centrally and shipped out, so there are no weights to refresh. When conditions change, the instance learns from live experience and adapts in operation. Nobody needs to schedule a retraining run.
Each instance runs on an ordinary processor, with no GPU or training infrastructure. A single CPU server supports hundreds of live streams, allowing multiple instances to keep learning while they operate.
INDUSTRIES
ADAPTIVE INTELLIGENCE
WORK WITH US
We welcome conversations with companies facing problems our AI could help solve, investors who share our ambition, and researchers who want to join our team and advance continual learning.
Contact usFAQs
