Integration view
Reference architectures
Three complete stacks, one row per stage: (a) a solo engineer on a low budget with open source, (b) a startup with a product vertical, (c) large-scale industry as publicly described. Use the columns to see what you can skip at your scale and what you will need when you grow.
What changes as you scale
- Data becomes the product. From (a) to (c) the share of effort moves from models to data operations: collection programs, labelling pipelines, catalogs, privacy and the flywheel.
- Evaluation becomes infrastructure. Solo teams run tens of trials; industry runs evaluation fleets, real2sim and world-model evaluators with measured correlation.
- Safety and regulation become schedule drivers. At (b) and (c), Machinery Regulation, CRA and AI Act dates shape release plans (stage 14).
- Compute splits. (a) runs inference on a desktop GPU; (b) on Jetson Thor with TensorRT; (c) adds learned fast layers (S0) and off-board reasoners.