Research in progress
What if a model learned
your way of working?
We’re researching actual custom models trained on a business’s knowledge, language and decisions, then testing how well they handle real work.
Can your way of thinking
become part of the model?
A company has more than documents. It has language, habits, exceptions and judgments. We want to understand which of these can be learned, what should stay explicit, and where people need to remain in control.
Open up the research.
Training, memory and retrieval are different.
Retrieval supplies a model with relevant information at the moment it answers. QIP memory carries context and lessons between tasks. Fine-tuning changes model parameters through training. We are investigating where each approach helps, and how they might work together.
Business-specific models, grounded in permission.
The training track explores adaptation using authorized examples of business language, decisions and work. It includes investigating whether fine-tuning is useful for a given problem. Training a foundation model from scratch is a separate feasibility question, not a promised offering.
Evaluate on work the model has not seen.
Held-out tasks, source-grounded answers and human review help test whether adaptation improves useful behavior. Baselines should include retrieval without training, so greater complexity has to earn its place.
Make room for results that say no.
A custom model will not be the right answer for every business. Research should identify limits, repeated mistakes and cases where simpler systems are sufficient.
Follow what we learn.
Get research updates as the questions become experiments and the experiments produce findings.
Get updates