Approach · ML Control Planes
Flux Pro Shop
Image Generation Control Plane
Overview
Production-grade image generation control plane built on HuggingFace Diffusers, optimized for Apple Silicon (MPS). It manages local GPU generation, remote API runners, and complex artifact lineage for full reproducibility.
By designing a unified Compile/Execute Intermediate Representation (IR), this architecture abstracts the generation paths for complex ML models (Flux, SDXL/Pony, and GGUF).
Technical Architecture
Core Engine & IR
- Python 3.12+ type-safe core
- Unified IR: WorkflowSpec → Compiler → ExecutionPlan → Runner → Result
- Backend abstraction (DiffusersBackend, SDXLBackend, ComfyUIBackend, ApiBackend)
Data & Lineage
- Content-addressed storage for immutable generation results
- Automated lineage tracking (source_run, derivation)
- JSON-based metadata-driven reproducibility
Why This Matters
Artifact Immutability
Mathematical guarantee of 100% artifact reproducibility through rigorous lineage tracking.
Performance Engineering
Backed by an incredibly fast test suite (4,600+ tests executing in under 20 seconds).
Hardware Abstraction
Dynamic routing capability between MPS (Metal Performance Shaders) and remote API execution environments.
Production Quality
Enterprise-grade structural resilience, bringing software engineering rigor to experimental ML pipelines.
Technical Depth
This project demonstrates understanding of:
- Distributed state management and coordination
- WebSocket protocols and real-time systems
- Serverless architecture patterns and edge computing
- WebAssembly compilation and optimization
- Production deployment and operational monitoring
- Type-safe system design across language boundaries
Development Context
Built as a learning platform to explore modern web infrastructure while creating something family members could actually use. The constraints—real users, need for reliability, performance requirements—drove architectural decisions and forced engagement with production concerns beyond local development.