Consistent output, every time
Extract the same fields from your documents, classify requests into your categories or write in a fixed structure, with fewer instructions and fewer retries per call.
Train and fine-tune language and voice models on your own data, when off-the-shelf models are not good enough for your language, your domain or your customers.
Fine-tune a hosted model, or train an open-weight one and run it where your data rules require.
neu Intelligence trains and fine-tunes language and voice models for businesses in Ireland and internationally. We prepare your data, train on a hosted or open-weight model, and measure the result against the tasks it has to do before it goes anywhere near your customers.
Fine-tuning is worth it when a model has to follow your format, tone or domain language every time, or when a smaller, cheaper model needs to match a larger one on one narrow job. If the problem is missing knowledge, retrieval over your own documents is usually the better first step, and we will tell you so.
Extract the same fields from your documents, classify requests into your categories or write in a fixed structure, with fewer instructions and fewer retries per call.
Train speech and text models for languages and accents that general tools handle poorly, so they sound and read the way your customers actually speak.
Fine-tune an open-weight model for one high-volume task, then run it on infrastructure you control, including EU hosting, instead of paying a large model per call.
Urdu is spoken by more than 200 million people, yet most voice tools we tried did not sound natural in it. So we trained our own: urdu-realtime-1, a realtime Urdu voice model built to sound like Urdu as people actually speak it, handling everyday conversations such as taking a food order or confirming an order. Trained on our own hardware.
Watch the demo on X ↗Define the task and how it will be judged. Collect and clean examples, remove personal data the task does not need, and build an evaluation set before any training starts.
Compare prompting, retrieval and fine-tuning on the same evaluation set. Train only where it clearly wins, and test for regressions on the cases the model must never get wrong.
Deploy behind your application with monitoring, then retrain on new examples on an agreed schedule, or hand over the data, scripts and model to your own team.
Fine-tune when the gap is behaviour: a format, a tone, a classification or a narrow skill the model gets wrong even with good instructions and examples. If the gap is knowledge that changes, such as prices, policies or stock, use retrieval. We test both on the same examples before recommending either.
It depends on the task. A narrow format or classification task can often improve with a few hundred good examples, while broader behaviour needs more. Quality matters more than volume: correct, consistent examples that represent the real work. We review what you have before quoting, and help you collect more where needed.
With an open-weight model, you can own the trained weights, subject to the base model’s licence. With a hosted provider’s fine-tuning, the model sits in your account on that platform, under its terms. We agree ownership, and where the training data is stored, in writing before we start.
Not by us. For hosted fine-tuning, check the provider’s current business terms, which generally exclude API data from training their own models by default. For sensitive data, an open-weight model on infrastructure you control keeps the data and the model out of third-party hands entirely.
Start with a free 30-minute conversation. We’ll look at the task, the tools you use and a practical first step.
Calendar not loading? Open it on Cal.com or email tash@neu.ie.