jepctl: Putting Joint-Embedding Predictive Architectures in Practitioners' Hands
jepctl is a single-binary local runtime written in Rust with no Python dependency that loads published JEPA-style checkpoints (I-JEPA, V-JEPA 2, DINOv2, AudioMAE), strictly validates weight coverage, and serves embeddings over REST, SSE, and CLI. This engineering report documents the systems architecture, runtime constraints, and single-machine performance profiles measured on an Apple Silicon (M5 Pro) testbench. It packages four worked integrations: few-shot inspection and recognition, an online latent world model predicting camera dynamics, an end-to-end robot arm controller driven by latent distance minimisation, and a multimodal audio-visual companion. Source code and models are open source (Apache-2.0): https://github.com/younss/jepctl
Authors
- Younss Azzayani
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-09-22
- DOI
- https://doi.org/10.5281/zenodo.22883995
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00