Circuit-JEPA

Goal — Build Schematic-Aware JEPA, a graph-based AI model that learns electronic-circuit structure through self-supervised pretraining and acts as a semantic co-pilot for an EDA tool, performing functional-block recognition and semantic design-rule checking. Components — Five layers: data sourcing (atopile registry, KiCad/Altium corpus, synthetic-negative generation), ingestion and canonicalization, graph representation, JEPA pretraining (context/target encoders, predictor), and EDA integration with equivalence-aware evaluation. Team — Roughly 6–8 people: a research lead, two graph-ML / self-supervised researchers, one or two data engineers, an electronics/EDA domain expert, an MLOps engineer, and a software engineer for tool integration. Skills required span graph neural networks, JEPA, Python data pipelines, circuit design and SPICE, and large-scale training infrastructure. Duration — Roughly 18–24 months to a validated prototype: about 6 months on data and the oracle, 9 on representation and pretraining, the remainder on downstream tasks, integration and evaluation.

No feedback yet. Be the first to share your thoughts!