Demo Build · Private AI · On-Prem RAG
A private AI document assistant that never sends data outside
Law firms, accounting offices, and clinics can't paste client documents into public AI chatbots. This assistant runs the language model on a server in their office — answers come from their own documents, with the source cited, and the internet can be switched off entirely.
Local LLM (Bielik)RAGOn-premSource citationsGDPR / EU AI Act
▶ 1-minute narrated demo — the internet is switched off mid-demo; every answer cites the exact document and clause. Ta oferta po polsku →
The challenge
Regulated businesses sit on thousands of documents — contracts, case files, procedures — and waste hours searching them. Cloud AI could help, but professional secrecy and GDPR make sending client data to third-party APIs a non-starter, and the EU AI Act's transparency obligations (in force since August 2026) raise the bar further. The result: the companies that need AI most are the ones least able to use it.
What I built
A document assistant where everything runs on hardware inside the company: a Polish-language open model (Bielik, Apache 2.0) served locally, retrieval over the firm's own document index, and a chat interface the whole team can use.
- Zero data egress: the model, the index, and every answer live on an on-prem server — the demo runs with networking switched off.
- Verifiable answers: each response cites the exact document and clause it drew from — one click to check.
- Knows its limits: when the answer isn't in the documents, it says so instead of guessing — no hallucinated policies.
- No metered costs: no per-token fees, no per-seat licenses; the server is a small, quiet box the size of a book.
How it works
- Documents (PDF/DOCX) are ingested into a local semantic index — nothing is uploaded anywhere.
- Each question retrieves the most relevant fragments; the local LLM answers from that context only.
- Sources — document and clause — are attached to every answer.
- Compliance layer for the EU AI Act: AI-interaction disclosure (art. 50), system documentation, and a usage register.
The impact
- Minutes of document-hunting become seconds — with the citation to prove the answer.
- GDPR exposure from cloud AI drops to zero: data never leaves the building.
- Predictable cost: hardware plus a maintenance plan — no per-token or per-user fees.
- Pilot live in 3–4 weeks, on the client's own documents.
Tech
Bielik 11B (Apache 2.0)MLX / llama.cppRetrieval-Augmented GenerationLocal embeddings (bge-m3)Mac mini / Mac Studio serverOpenAI-compatible local API
Shown here as a working demo for a fictional law office (“Nowicka & Partners”). The same system drops onto your real documents and hardware. Polish-language offer: progrise.dev/prywatne-ai.
Want this running on your documents?
A 30-minute demo — on documents from your industry, with the Wi-Fi off. No commitment.
Get in touch →