Models trained on your work, running on your hardware.
General models are a good start. Models tuned on your own decisions are better at your own work. Nexus schedules training on your graphics cards, tests every model before it serves anyone, and watches for drift once it does.
Tuned on your decisions, with data you are allowed to use.
Fine-tune base models on your own records, with the datasets checked for what may and may not be used. Compare runs side by side before choosing one.
Fine-tuning runs
- Runs compared side by side
- Base models from your own store
- Cost per run
Datasets
- Consent and purpose checked
- Personal data found first
- Refusals explained
Trained on decisions that held up
Train a model only on sealed decisions that stood up to review.
- Only decisions that held up
- Linked back to their records
- Kept inside your boundary
No model serves anyone until it passes.
Every model is evaluated against your own tests before it is registered. Models that fail are refused, with the reason. Those that pass go live gradually, a share of requests at a time.
Evaluations
- Your own test sets
- Pass marks per task
- Results kept with the model
Deploy a model
- Staged by share of requests
- Compared with the model it replaces
- Back in one step
When a model starts getting it wrong, you know.
Nexus watches every model in use. When answers start to drift from what people decide, you see where and why, and can start retraining from the same screen.
- Answers compared with people’s decisions
- Drift shown by task and unit
- Retraining started from the alert
About your own models.
Do we need our own graphics cards?
No. Hosted models work from day one. Training your own needs graphics cards, in your data centre or in ours.
Can training data leave our site?
Not unless you allow it. You set where a job may run for each dataset, and anything that could send data out is checked first.
Go deeper
Tune a model on your own decisions.
Bring one task your team decides every day. We will show what a tuned model does with it.