Hranitel — AI document archive
A closed-loop AI system for law firms and document-heavy businesses: it reads scans on its own and pulls out the data you need, a document is found by meaning in seconds, and files never leave the company perimeter.
Challenge
At a law firm with a multi-year archive, finding one contract turns into digging through folders full of files named "Contract_final_2.pdf". Who sued whom, on what terms, and when — lives in one lawyer's head and leaves with them when they quit. Uploading documents to public AI tools for search isn't an option: it's someone else's trade secret.
Solution
Hranitel runs inside the client's closed loop — on their own server or an isolated GPU cloud in Russia, with no data leaving the perimeter. The system reads scans and photos of documents on its own, finds people, organizations, sums and legal references in the text — and links each file to the right case. Documents with shared entities cross-link automatically, Wikipedia-style, and the whole archive can be read through 6 structures at once: by client, by type, by timeline, by status, by participant, and a "Forest" view where the client is a tree and the document is a leaf. An AI chat answers questions with a direct quote and a link to the source page.
Outcome
The product is at the commercial-proposal stage (v2.1): the frontend already runs on a demo law-firm archive, part of the backend logic is stubbed for a specific rollout. By the vendor's numbers: finding a contract — 2 seconds instead of 15–30 minutes, parsing a scan — automatic instead of 5–10 minutes by hand, pulling a case file for court — 15 minutes instead of two days.
How a document enters the system
- 01
Upload
A phone scan, a PDF from email, or a photo from a site visit — filed manually as client → case → type, or dropped into an "Unsorted" bucket.
- 02
Reading the document
The system reads the scanned text itself — Russian or English, without sending the file anywhere else.
- 03
Understanding the content
The program finds companies, people, courts, sums and legal references in the text — and links the document to the right case on its own.
- 04
Cross-linking
Documents sharing entities link automatically, plus explicit relations like "reply to claim" or "appeal against ruling".
- 05
6 structure models
One archive, read differently: clients, types, timeline, status kanban, participants, and "Forest".
- 06
AI search
The chat answers plain-language questions with a direct quote and a link to the source page.
- 07
Access
Owner / Admin / Employee roles — everyone sees only what they're supposed to.
Delivery and economics
Client's own server — a one-time ₽350,000, 12GB+ GPU and 32GB+ RAM. Alba Creation's Russian cloud — ₽25,000/mo plus a one-time ₽50,000 setup, live in a week.