AI on your data, without your data leaving
Some organizations cannot use a public AI API. Not by preference: by law. Clinics, law firms, notaries, financial institutions and public bodies handle information that cannot leave their premises.
There is a third way: run and train the models on your own infrastructure.
Three capabilities that combine
1. Training models on your data
QLoRA fine-tuning of open models, trained with your internal documentation. We have empirically characterized the limits of this technique on an NVIDIA DGX Spark workstation with 121.7 GB of unified memory, training six configurations from 3 to 72 billion parameters. The work is published as a technical report.
2. RAG engines over your documentation
Retrieval-augmented generation: the system indexes your manuals, contracts or records and answers specific questions citing the source document. When it finds no support in your documents, it says so instead of making things up. The whole chain — embeddings, vector store and answer model — can run inside your network.
3. Connecting agents via MCP
The Model Context Protocol is the open standard that lets a model use your systems as tools: query the ERP, open a ticket, read a repository or trigger a process. We build the MCP servers that expose your systems in a controlled way, with per-tool permissions and full traceability.
Measured, not estimated
Real peak memory in QLoRA fine-tuning on GB10 (Grace-Blackwell, 121.7 GB unified):
| Model | Type | Peak VRAM | Peak RAM | Time |
|---|---|---|---|---|
| Qwen2.5-3B | Dense 4-bit | 2.66 GB | 13 GB | 15.4 s |
| Qwen2.5-7B | Dense 4-bit | 8.53 GB | 29 GB | 19.7 s |
| Qwen2.5-14B | Dense 4-bit | 18.53 GB | 51 GB | 35.1 s |
| Qwen3-30B-A3B | MoE bf16 | 72.02 GB | 95 GB | 518 s |
| Qwen2.5-72B | Dense 4-bit | 43.96 GB | 64 GB | 24 s |
Sequence length 256–4096 depending on configuration. Figures come from our R&D test bench on a reduced dataset — they size the hardware envelope; they are not a performance promise.
What we deliver
- MLOps suite deployed on your infrastructure — Docker container with REST API and a real-time memory telemetry dashboard.
- Heterogeneous data ingestion — Excel, Word, PDF, Markdown, CSV and TXT automatically converted to trainable JSONL.
- QLoRA/SFT training engine with memory guards and a host watchdog.
- Custom MCP servers with per-tool permissions and full call logging.
- Hardware sizing — what you actually need for your model size and context length.
When it makes sense
When regulation stops you from sending data to a public API. When per-token costs no longer beat running locally. When you need a model that speaks your industry’s vocabulary. Or when you want an agent operating over internal systems with auditable permissions.