Problem

Legal work is spread across official proceedings, documents hundreds of pages long and deadlines that must not be missed. Files and papers live in different places, understanding a large file takes time, and a missed deadline is a serious professional risk. An AI answer without a clear source is unusable in law, and a firm's knowledge often stays with individuals.

Approach

One workspace. Legal matters, official proceedings, documents, parties, hearings, deadlines and tasks share one data model. The screens of UYAP (Türkiye's national judiciary information system) are not rebuilt; UYAP is connected as one read-only source.

An immutable document platform. Every uploaded document is stored as a version; malware scanning, text extraction and OCR, classification and indexing run in the background. The original is never altered.

Cited question answering. The assistant searches only within what the user is authorised to see, and every claim is linked to a document version, page and text span. When the evidence is insufficient, it says so.

Deadlines through a rule engine. A deadline proposal is calculated from an event such as service of a judgment using a versioned rule; for critical deadlines, the trainee's check and the lawyer's approval are separate steps. An approved deadline is linked to tasks, reminders and the calendar.

Controlled drafting of court submissions. A drafting flow built on source selection, a template, an argument plan and evidence matching, with every step open to the lawyer's review.

Critical decisions

AI proposes, the lawyer decides

A proposed date, party, amount or deadline never becomes a confirmed record, timeline entry or calendar event without the user's approval. The same person cannot complete both approval steps.

No answer without a source

Every citation is linked to the document version and text span used. When a new version or a new event arrives, the earlier analysis is marked as no longer valid.

Permissions before search

Ethical walls and permission filters are applied before the search runs. A file a user cannot see does not appear in results, counts or suggestions; even its existence is not revealed.

Official, user and AI data kept apart

Official sources, user corrections and AI-derived content never overwrite each other and are shown separately on screen. Users can see why a result was produced.

Data stays in Türkiye

The application, documents, indexes, models, logs and backups are kept in Türkiye. UYAP passwords and e-signature PINs are never sent to the server.

Deliberate limits

The first version does not file documents to UYAP, does not complete official actions with e-signatures and does not predict case outcomes. The platform speeds up the lawyer's work; it does not replace it.

Details

An example flow

A matter is opened in the workspace for a lease dispute, and the first-instance and appeal files are linked. When the reasoned judgment is uploaded, its text is extracted and indexed. The assistant proposes the date of service and links that proposal to the relevant page of the judgment. A trainee checks the proposal and the lawyer approves it; only then is the deadline confirmed and preparation tasks and reminders proposed. Every step is recorded with who did it, when and from which source.

This scenario describes the product flow; it does not come from a real case.

Scope

Identity, workspaces, roles and ethical walls
Legal matters, official proceedings, parties and representation
PDF, DOCX, UDF and image documents with versions, text extraction and source location
Cited question answering, summaries and timelines
Deadline rules, hearings, tasks and reminders
Petition drafts with templates, writing profiles and the firm's knowledge base
Retention, deletion, restore and audit trail

Development stages

Stage	Exit condition
Core platform	Workspace isolation and audit tests pass.
Document core	Source location and processing state are visible to the user.
Cited AI	Cited answers are produced without permission leakage.
Timeline and deadlines	Rule versions and the approval chain are traceable.
Workflow and integration	Partial sync and channel failures are handled.
Drafts and knowledge base	Cited drafting and the lawyer's revision flow work.
Desktop and hardening	Pilot security and restore acceptance are complete.

Technical design

A modular core application, separate Python workers for heavy jobs (OCR, classification, indexing, model inference), PostgreSQL with row-level security and vector search, an immutable versioned object store and queue-based background jobs. Web, PWA and a thin desktop client, with multi-factor authentication. Language, embedding and reranking models are served within Türkiye. Long jobs never block the user's request, and every critical action is audited with who performed it and from which source.