Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models

Kairos addresses a critical bottleneck in current Time Series Foundation Models (TSFMs): the inability of static tokenization schemes (like those in Chronos or Moirai) to handle irregularly sampled data or varying periodicities without massive parameter overhead. While technically sound and addressing a legitimate research gap, the project currently lacks any significant market traction (0 stars, though 7 forks indicate some academic activity). The defensibility is low because the 'moat' is purely algorithmic; if the proposed adaptive tokenization proves superior, it will likely be absorbed as an architectural refinement into dominant industry models like Amazon's Chronos, Google's TimesFM, or Salesforce's Moirai. Frontier labs and cloud providers have a high incentive to dominate the time series space as it is central to enterprise AI (demand forecasting, anomaly detection). Consequently, Kairos faces a high risk of 'feature-ization' by larger platforms within the next 12-24 months.