A corporate AI platform that processes documents, answers with citations, identifies contractual risks and turns each transaction into a manageable project.
Contracts, acceptance certificates, invoices and policies are scattered across folders, email and messaging apps. Search depends on filenames and employee memory, and a generic AI answer is difficult to use for legal work without a page reference and quotation. We created a connected workflow taking documents from secure upload and OCR through search, risk analysis, tasks and closing documents.
Employees search for files by name or ask a colleague who remembers where the correct version is stored.
We organised documents by workspace, folder, collection and tag, and added hybrid keyword and semantic search.
Users see a polished answer but cannot tell which document supports it.
The RAG chat answers only from the available context and returns the document, page, quotation and relevant passage.
Legal and financial terms still have to be checked manually immediately before signing.
AI review classifies each issue and displays the original quotation, recommendation, severity and processing status.
Amounts, obligations, deadlines and completion documents are still managed in spreadsheets.
We turned a set of documents into a transaction project with progress tracking, owners, deadlines, risks and a clear next action.
The architecture was verified using the supplied documentation, application structure and genuine product screens.
A Celery worker extracts text from PDF, DOCX, XLSX, PPTX, HTML and images, applies OCR for RU/KZ/EN, classifies files and builds the index. Employees receive streaming chat answers with sources, while project views show managers amounts, obligations, risks, deadlines and the next action.
Dedicated members, roles, documents and Qdrant collection
Original in MinIO, metadata in PostgreSQL, chunks in Qdrant
Severity, quote, recommendation and workflow status
Parties, amount, obligations, deadlines, tasks and supporting files
Variables, instructions, AI-assisted completion and DOCX/MD/HTML export
Organizations, BIN/IIN, contracts, completion certificates and document relationships
The response appears progressively without waiting for the full LLM cycle
Every conclusion is linked to a document, page and excerpt
In progress, resolved, risk accepted or not applicable
Instructions, retrieval depth, temperature and legal mode
Security is layered: access JWTs expire after 15 minutes, and rotating refresh tokens last 30 days in an httpOnly cookie. Reuse of an old refresh token revokes the related token family. The backend verifies workspace roles, APIs use CSRF protection, rate limiting and security headers, and uploads are checked by actual MIME type and SHA-256. Auditing records document, template, agent and access operations.
The current stage is a functional MVP with an architecture designed for production. Client deployment requires final CI/E2E checks, infrastructure configuration and acceptance testing. Demonstration data on the screens is not presented as real client documents.
Global user/admin roles are combined with an owner → admin → editor → viewer hierarchy inside each workspace. Activity logs and dedicated workspace settings give managers a controlled collaboration model, while the backend remains the authoritative source for permission checks.
Qdrant and PostgreSQL FTS are combined through Reciprocal Rank Fusion.
FastEmbed computes embeddings on the CPU without sending files to an external service.
Celery separates intensive OCR and parsing from the user request.
Structured logs, Prometheus, Sentry and health checks.
Alembic, pytest, Vitest, Playwright, Ruff and TypeScript typecheck.
Docker Compose, Traefik 3.6, HTTPS/Let’s Encrypt and backup/restore scripts.
The platform sells a managed process—from file upload to decision, task and final document—not an “AI chat”.
We will analyse the audiences, content, integrations and search demand, then design the structure and technical architecture before visual design begins.
Discuss the Project ↗Back to Case Studies