PDF to DSTV CNC reconstruction

Five-phase R&D pilot with a steel-fabrication company. The research question: can a working prototype reconstruct CNC programs from 2D PDF fabrication drawings with enough reliability to exit the sandbox and justify a productisation decision?

The approach combines supervised fine-tuning for format discipline, reinforcement learning with multiple reward components on a progressive curriculum, and a safety architecture with uncertainty signaling, out-of-distribution detection, and multi-view consistency checks. Ablation studies validate each component.

The team mirrors how we run delivery: project lead from Step into Liquid, ML consulting from Duco (NKI/TU Delft), HvA research supervision from Jurjen Helmus with an intern on RL algorithm comparison, and development at Social Technology Lab with Paul and Khaled. HGG cofinanciers and supplies the data; TechValley coordinates the KvW3/EFRO subsidy administration.

All phases include a Go/No-Go gate. The KPIs are research targets, not performance guarantees. This is the kind of environment the training is built for: where shallow AI judgment costs real money.