EdgeOrchestra

On top of the infrastructure, a research project: a federated instantiation of LeJEPA — Balestriero & LeCun's heuristics-free self-supervised JEPA. Reproduced the centralized reference behavior (88.5% linear probe on CIFAR-10), then moved it to severe non-IID federated settings (Dirichlet partitions) with plain FedAvg.

Along the way, identified and diagnosed a critical failure mode of the implementation — collapse caused by an under-weighted Epps-Pulley statistic — pinned down through exact numerical fingerprinting.

Research-grade methodology: pre-registered decision rules written before the numbers, bit-exact reproducibility (seeded runs, per-run config + commit provenance), an adversarial self-review protocol, and a 30+ entry lab notebook.

Open-source federated learning infrastructure that transforms unused Apple devices (iPhone, iPad, Mac) into a distributed ML compute cluster. Data stays on each device — only gradients are shared, preserving privacy.

The orchestrator (FastAPI + gRPC) handles device registration, training coordination (FedAvg), model registry with CoreML model building, and battery/thermal-aware scheduling. Backed by PostgreSQL and Redis, with a web dashboard (HTMX + Chart.js) and Prometheus metrics.

Native iOS/macOS workers built with SwiftUI perform real on-device training via CoreML MLUpdateTask. Gradients are quantized Float32→Float16 and compressed with LZ4 before transmission. Zero-config discovery via Bonjour/mDNS.

The model registry supports built-in architectures (MNIST, CIFAR-10) and custom models via declarative specs. Server-side evaluation runs numpy forward passes on held-out test sets to track convergence.

Includes a Python CLI control plane, a worker simulator for testing without hardware, TLS/mTLS support, and a complete Docker Compose deployment. Targeting academic publication at MLSys, MobiCom, and NeurIPS FL Workshop.