The full answer
PyTorch is Creuto's usual choice for custom model development and experimentation, while TensorFlow is used where its deployment tooling or an existing client model makes it the better fit. Both are applied to forecasting, classification, anomaly detection, recommendation and computer vision.
Hugging Face provides open models and the libraries to run them. Creuto uses it to evaluate open-weight language models such as Llama and Mistral, pull embedding models for retrieval, and fine-tune smaller task-specific models when a commercial API is too costly or data cannot leave your environment.
Not every problem needs a trained model. A rules-based approach or a prompted LLM often reaches the goal faster, and Creuto compares those options against a simple baseline before investing in training.
Models that reach production are packaged in Docker containers, served through FastAPI, monitored for drift and retrained on a schedule, with the pipeline defined in code so your team can run it after handover.
If you have data scientists with notebooks that never reached production, Creuto's MLOps work turns those experiments into reliable services. Aquapulse, the aquaculture platform used by 6,000+ farmers, shows how data-driven forecasting can reach users in the field.