Open-weight models, local
DeepSeek V4-Flash, Llama, Qwen — running on the appliance's GPUs.
UBOTIKA
On-premises · Air-gapped · Managed
A private AI system that never leaves your office.
Firm-owned hardware, installed in your office. No internet connection. No third-party model provider. The models run on the appliance; your documents and query history stay on storage the firm owns.
- Outbound network
- None
- Access path
- Firm LAN only
- Commitment
- Month-to-month
- Install window
- [X] weeks
01 / The problem
Where the confidentiality question sits
- Client data sent to a commercial AI provider leaves the firm's control — governed by someone else's systems and terms.
- Retention windows, training-use carve-outs, and subprocessor lists vary between providers — and they change.
- Some clients and outside-counsel guidelines now restrict third-party AI use on their matters.
We make no claim about how courts have treated any of this. An on-premises system removes the third-party transfer from the analysis.
02 / What we install
One appliance, on your premises
Private web UI, LAN
A browser on the firm network. No client software, no accounts with us.
No outbound path
Verifiable at your firewall and on the box during the pilot.
Search your own files
We index the shares you designate; indexes stay on firm storage.
Footprint [X]U · sized for [X] concurrent users · firm provides rack space, power, LAN port, storage.
03 / Data separation architecture
Box A — UBOTIKA-owned
UBOTIKA appliance
Model weights · inference software · GPU hardware, OS, serving stack · no document or query storage.
Box B — Firm-owned
Firm-owned storage
All documents · all query history · all search indexes · encrypted at rest, firm holds the keys.
If the engagement ends, the appliance is removed and firm data remains entirely with the firm — documents, history, and indexes are already on your storage, so there is nothing to migrate back.
04 / The subscription
A managed service, month-to-month
- Hardware, sized and replaced by us
- On-site installation with your IT
- Model deployment and indexing
- Updates, on your approval
- Monitoring — metrics only
- Support, [X]-hour response
No hardware purchase and no capital request. One monthly fee, scoped to the appliance your firm needs.
05 / Honest limitations
- Local open-weight models are strong but not identical to frontier cloud models; the hardest reasoning and long-context tasks show it.
- Throughput is sized to concurrent users. We scope hardware to firm size; heavy simultaneous use queues rather than scales on demand.
- This does not replace the firm's document management system. Your DMS remains the system of record.
- It is not a substitute for legal review of AI output. Every result still needs a lawyer's judgment.
06 / Pilot
Evaluate it on your own network first
A [X]-week pilot on one appliance with [X] attorneys and your IT director, who verifies the isolation. If it does not hold up, we remove the appliance and the engagement ends.
- Your IT director verifies there is no outbound path, on your firewall
- Documents and query history stay on firm storage throughout
- Month-to-month after the pilot — no capital request