The CBCT scan validator at Hybridge is the project that taught me what production ML actually costs. The model is the small part: a self-supervised spatiotemporal masked autoencoder that runs inference in 58 milliseconds on half a gigabyte of VRAM. The big part is everything that keeps it honest: a fully automated, versioned training data pipeline where training and inference share one preprocessing path, and a weekly CI/CT job that evaluates against a deterministic holdout and refuses to promote a model that fails the gate.
I benchmarked six architectures behind one shared interface before picking the production model: a 2D ConvNeXt baseline, a hybrid transformer, a 3D CNN, an nnU-Net adapter, a VAE anomaly detector, and the ST-MAE that won. When the model is wrong, confidence thresholds route the scan to a human instead of pretending.
Before Hybridge, I spent a year and a half at NYU researching whether RLHF methods (PPO, GRPO) could make Llama a measurably better persuader. Reward modeling and preference optimization, the same machinery behind modern aligned models, pointed at argument quality. That training-side depth is why I trust myself to debug the serving side.
