A fresh approach. Coming soon.
Good infrastructure.
Thoughtful models.
People who know the work.
We’re building a hosting company we’d want to use ourselves. A home for your projects, with custom model training backed by years of hands-on experience.
Tell us what you’re buildingWhat’s taking shape
A straightforward home for your applications, development environments, and self-hosted tools. We’re building VM hosting with clear resources and a focused console, so you can get on with your work.
Coming soonYour data. Your problem. A model built around both. We’re bringing 15 years of statistical analysis and model-training experience to custom training, with engineers who understand the work behind the results.
Coming soonExperience before the buzzwords
The language has changed. Our engineering mindset hasn’t. We’ve spent 15 years working with data and training models. Now we’re bringing that experience to Fresh.
Practical ML engineering
We build and train machine learning systems against real-world datasets at serious scale.
Our experience includes product classification and data enrichment systems developed for Flywheel and Data Purl, ranging from transformer models trained across tens of millions of ecommerce records to hierarchical classifiers trained on more than 3.5 million labelled products.
We’ve fine-tuned transformer architectures including BERT, DistilBERT and Gemma, built custom multi-head neural networks, and developed taxonomy-aware training and inference systems capable of classifying products through complex multi-level category structures while guaranteeing valid parent-child relationships.
Our work goes beyond simply fine-tuning a model. We design the full training pipeline: data preparation, embeddings, sampling strategies, class balancing, custom loss functions, LoRA/QLoRA fine-tuning, evaluation, confidence calibration and high-throughput inference.
We’ve also developed semantic embedding and clustering systems for product matching and canonicalisation, combining automated modelling with human feedback to create continuously improving training datasets.
The result is practical ML engineering designed for production — models that can be trained, evaluated, versioned and deployed to process millions of records reliably and economically.
Let’s start a conversation
Have a hosting project in mind, or a model you want to train?
We’d like to hear about it while we build what’s next.