How Does EdgeCraft Turn Intent Into Deployable Edge ML Models?

The problem it solves

Producing a deployable edge ML artifact for a specific scenario requires navigating a huge search space. That space spans data representation, model design, training on domain-specific data, and runtime customization. The workflow is fragmented and difficult to scale across diverse edge applications. EdgeCraft addresses this by turning high-level intent into deployable edge ML artifacts using an LLM-driven approach.

Two challenges and two designs

Building such a system raises two challenges. First, how can an LLM be guided to find high-quality solutions that meet dynamic SLOs for task quality, latency, and energy? Second, how can trustworthy target-device verification be obtained at low cost? EdgeCraft answers with two designs. A constraint-aware synthesis tree explores alternative candidates and uses measured SLO gaps to guide each improvement. A multi-fidelity verifier progressively combines low-cost checks with full target-device verification to reduce verification cost while preserving reliable verification results.

Verification that learns from failure

The verifier also records verified failures for reuse. This avoids repeated device work. It progressively combines low-cost checks with full target-device verification. The combination reduces verification cost while preserving reliable verification results. The paper does not specify how many failure records are stored or for how long.

Concurrency through multi-tenant runtime

To support concurrency, EdgeCraft provides a multi-tenant runtime. It runs cloud training and target-device verification in parallel while isolating requests. The document does not detail the isolation mechanism or the maximum number of concurrent tenants.

Results across 50 public tasks

Across 50 public tasks, EdgeCraft exceeds the task-specific Reference in best-observed quality on 40 tasks. It finds an SLO-feasible artifact on 45 tasks. The two outcomes overlap on 38 tasks. Moreover, EdgeCraft achieves competitive performance on a self-collected SEN dataset. This suggests generalizability to real-world IoT sensing tasks. The paper does not report latency or energy numbers for the SEN dataset.

What the paper leaves open

The paper spans 20 pages, 15 figures, and 5 tables. It does not release code or model weights. It does not describe the LLM backbone used. It does not quantify the verification cost reduction in absolute terms. It does not compare against other automated ML systems by name.