The Sovereign AI Flywheel
Your Team Uses the Tool.
The Tool Becomes Your AI.
CaveauAI isn't a search engine you rent. It's a data engine that builds a proprietary AI model from your daily workflows — entirely on European infrastructure you control.
Instant Utility — Day One
Your team uploads their working documents — contracts, case law, internal manuals, regulatory filings — to our EU-hosted bare-metal servers. Within minutes, they have enterprise-grade search across everything they uploaded.
No training required. No AI expertise needed. They type a question in plain language, and CaveauAI returns precise answers with clickable citations back to the source document. Powered by open-source models up to 72 billion parameters.
- Upload PDF, DOCX, TXT, HTML, or Markdown — up to 50 MB per file
- Vector, keyword, and hybrid search modes
- Every answer cites its source document and paragraph
- EU-hosted infrastructure — your data never crosses jurisdictional boundaries
The Invisible Asset Build
Here is where CaveauAI diverges from every other enterprise search tool on the market.
While your team does their normal work — asking questions, clicking cited sources, accepting or refining answers — the platform is quietly building something valuable in the background. Every interaction is securely logged on your isolated database: the question asked, the context retrieved, the answer generated, and whether your expert accepted it.
By simply using the tool, your organisation is passively generating a high-quality, human-verified, domain-specific instruction dataset. No annotation team. No data labelling project. No extra budget line. Your daily workflows produce the training data.
- Every query-answer pair is logged with full context
- Expert acceptance signals act as implicit human verification
- Data stays in your isolated database — never shared, never aggregated
- Zero additional effort from your team
Your Sovereign Model — Month 4+
Once your organisation has generated enough high-quality interactions — typically four to six months of normal use — we take that proprietary dataset and fine-tune a compact, specialised model exclusively for your business.
This model runs on our dedicated NVIDIA RTX PRO 6000 infrastructure. It understands your corporate terminology, your document structures, your regulatory context, and your team's specific way of asking questions. It doesn't guess — it knows.
- Fine-tuned on your verified interaction data — not generic internet text
- Smaller model, drastically lower inference cost, faster responses
- Runs on dedicated GPU hardware — not shared cloud instances
- The model belongs to your organisation
What This Means for Your Organisation
No Vendor Lock-in
You aren't renting an AI service. You are building a proprietary asset that understands your corporate language, your contract structures, your regulatory environment. If you leave, the model and the data go with you. Your investment compounds — it doesn't evaporate when a subscription ends.
Cost & Speed
A fine-tuned 7B parameter model that knows your domain outperforms a generic 70B model on your specific tasks — at a fraction of the compute cost. Smaller model, faster inference, lower monthly spend. The economics improve the longer you use the platform.
Absolute Data Sovereignty
The entire process — from initial document upload, through every query your team runs, to the final fine-tuning of your custom model — happens on our secure European bare-metal servers. Your data never touches OpenAI, Google, or any third-party AI provider. Full GDPR compliance by architecture, not by policy.
Built on Hardware You Can Point To
No abstraction layers. No "somewhere in the cloud." Dedicated European bare-metal servers with GPUs allocated to your workload.
Your Data Never Leaves Your Control
Cryptographically isolated MariaDB per tenant. 10.8.0.x WireGuard mesh encryption. Dedicated bare-metal NVIDIA RTX Blackwell clusters. No data ever traverses the public internet or touches a US cloud API. You own the GGUF model artifact.
No Ambient Telemetry
We log exactly what you permit to build your model, and nothing more. No tracking pixels, no shared tenant tables. Full GDPR compliance by physical architecture, not just by policy.
Start Phase 1 today. Upload your first document for free →
Generic AI vs. Your Sovereign Model
| ChatGPT / CopilotRenting | CaveauAI FlywheelOwning | |
|---|---|---|
| Data location | US cloud (OpenAI / Microsoft) | EU bare-metal (your tenant) |
| Model ownership | Vendor-owned, shared | Your organisation owns it |
| Domain knowledge | Generic, internet-trained | Fine-tuned on your verified data |
| Training data used | Your prompts may train their model | Your data trains only your model |
| Cost trajectory | Fixed per-seat, no compounding | Decreasing — smaller model, lower cost |
| If you cancel | You keep nothing | You keep the model and data |
How It Actually Works
For the security team. No marketing. Just the architecture.
Data Logging
We log what you explicitly permit. Nothing more.
When a user submits a query, the platform records four things: the raw query string, the retrieved document chunks (context window), the generated response, and the user's feedback signal (accept, reject, or refine). Each record is timestamped, tied to a session ID, and written to your isolated MariaDB instance — a dedicated database on your tenant, not a shared table.
There is no ambient telemetry. We do not log mouse movements, session duration, navigation patterns, or anything outside the explicit query-response-feedback loop. The logging schema is documented and available for audit. You can disable logging entirely at the tenant level, though this removes your ability to accumulate fine-tuning data.
The Air-Gap
The fine-tuning process does not happen on the same servers that handle your daily queries. It runs on a dedicated GPU cluster code-named hippo — physically separate Hetzner bare-metal nodes provisioned exclusively for training workloads.
Data transfer between your CaveauAI tenant and the training cluster happens over our private WireGuard mesh VPN on the 10.8.0.x subnet. No data traverses the public internet. The WireGuard tunnel is point-to-point with pre-shared keys rotated on a defined schedule.
At no point in this pipeline does your data touch a cloud provider's API, a shared compute cluster, or a network segment accessible from the public internet. The training data is purged from hippo after successful model delivery and client confirmation.
The Output
The deliverable is a set of model weights. Specifically: a quantised GGUF file (or safetensors if you prefer full-precision) containing your fine-tuned 8B parameter model, plus the LoRA adapter weights used during training.
This file belongs entirely to your organisation. It is not a license. It is not a service subscription. It is a binary artefact that encodes your corporate knowledge, tuned on your verified interaction data, that you can run on any compatible inference engine — llama.cpp, vLLM, Ollama, or our managed infrastructure.
We do not retain a copy of your weights. We do not use your model to serve other clients. We do not use your training data to improve a shared model. After delivery and your written confirmation, the weights and training artefacts are purged from our training infrastructure. You receive the only copy.
Start Building Your AI Asset Today
The sooner your team starts using CaveauAI, the sooner you accumulate the data that makes your sovereign model possible. Day one delivers instant value. Month four delivers a competitive moat.