Boudica Intelligence
OmniIndex
April 13, 2026
The Sovereign AI Newsletter · OmniIndex
Three engineering breakthroughs that make Boudica Torc more powerful, more flexible, and, most importantly — still entirely yours
One of the things we have always believed at OmniIndex is that a single monolithic model is the wrong answer for enterprise AI. The real world is made up of very different tasks — log analysis, contract review, customer data querying, report generation — and no single set of weights does all of them brilliantly.
So we built something better. Boudica Torc now supports multi-modal hot switching: you can run many specialist micro-models simultaneously, each between 500M and 3B parameters, each laser-focused on a specific workload. The Boudica Cognitive AI engine watches what you ask and routes the request to whichever specialist is most appropriate — automatically, with no input from you.
We know that some open-source general models — like Google’s Gemma — have capabilities built up from vast training corpora that would take considerable time to replicate from scratch. So rather than ask you to choose between sovereignty and breadth, we have done something pragmatic:
Boudica Torc now loads external open-source models and manages them as first-class citizens alongside your native Boudica specialists.
You can load one external model at a time. It participates in hot and warm switching exactly like a native model. Critically — and this is the part that matters — it operates under complete Boudica control:
The result: you get the general-purpose breadth of a large public model, delivered inside your sovereign boundary, without Shadow AI risk.
This one took considerable work. In standard practice, LoRA adapters and RAG data are tied to the model architecture they were trained against. A Boudica adapter works on a Boudica model; a Gemma adapter works on Gemma. The two worlds do not mix.
Our engineering team has changed that. Both LoRA adapters and RAG data created with Boudica Torc are now universally available — they can be applied to every model in your deployment, whether that is a native Boudica specialist or an external open-source model like Gemma. Train once, apply everywhere.
Here is a prompt that illustrates how these three features combine in a single request:
The Cognitive AI engine decomposes this request and allocates each part to the right resource. The external model handles HTML formatting — something it is already excellent at from its training — without you having to build a dedicated formatting specialist. That slot in your hot-loaded model pool stays free for something genuinely unique to your business, like your log analysis model.
This is the philosophy behind Boudica Torc: you should not have to choose between sovereign and capable. You should not have to rebuild what already exists. You should be able to focus your training investment on the knowledge only you possess, and let the platform handle the rest.
We are genuinely excited about what these features unlock, and we have only scratched the surface of the workflows they make possible. As always, if you would like a demonstration or want to discuss how these capabilities fit your specific environment, reach out to the team at info@omniindex.io.
The future of enterprise AI is on-premises, transparent, and sovereign — and it just got considerably smarter. Look out for additional customer stories over the next week or two, to see how they are using this new technology.