PLAIground: AI model switching system hits 91.3% accuracy across edge-cloud-space
Dynamic model selection prevents 21x cost violations in compound AI systems.
PLAIground addresses a critical challenge in the 3D Computing Continuum (edge, cloud, space): how to dynamically select the best AI model for each subtask in a compound AI system without breaking the application workflow. Traditional approaches either use a fixed model, which fails to adapt to changing conditions, or require manual reconfiguration. PLAIground introduces the Compoundable AI Model (CAIM) abstraction, which separates the 'what' (task semantics) from the 'how' (model implementation) using formal Task and Data Contracts. This means developers can swap models at runtime without rewriting workflow logic.
The framework's core innovation is the Pixie algorithm—an SLO-driven runtime model selector that continuously evaluates candidate models based on accuracy, latency, and cost trade-offs. In experiments with two realistic compound AI workflows (e.g., object detection + time-series analytics), Pixie achieved up to 91.3% accuracy while strictly adhering to SLOs. In contrast, fixed-model strategies either violated cost/latency budgets by as much as 21x or missed accuracy targets by 4%. This work, available on arXiv, is a step toward reliable, autonomous AI systems in distributed and resource-constrained environments.
- PLAIground uses CAIM abstraction with Task and Data Contracts to decouple model selection from workflow logic.
- Pixie algorithm achieves 91.3% accuracy while maintaining SLO compliance on cost and latency.
- Fixed-model strategies in comparison violate budgets up to 21x or miss accuracy targets by 4%.
Why It Matters
Enables reliable, cost-efficient AI orchestration across edge, cloud, and space—critical for autonomous systems and IoT.