Architecture

A multi-agent system that stops for human sign-off

Multiple specialized AI agents with tool access to the matter file, retrieval-augmented generation with exact citations, tenant-isolated indexes, and an approval layer binding every write operation to an explicit attorney decision.

  • Orchestration

    Specialized agents, unified data space

    Reception, Scheduling, Matter Work, and Deadlines operate as distinct agents with tailored toolkits. Handovers occur not over an unpredictable chat bus, but through the shared, verified state of client and case records: what reception records becomes immediate context for scheduling.

  • Human-in-the-Loop

    Cryptographically signed tool approvals

    Read tools execute fluidly; write operations never do. Every write proposal is sent to the client interface with a signed cryptographic token and strictly validated server-side before execution. If the signature secret is missing, the assistant fails closed rather than executing unverified mutations.

  • RAG

    Answers with verifiable citations

    Documents, file notes, appointments, and deadlines reside in a tenant-isolated PostgreSQL vector index (pgvector). Answers cite verifiable source passages, including page numbers whenever available in source PDFs. Juna leverages the same case context for pleading drafts and party captions. Scanned documents undergo OCR, with repair queues for failed scans.

  • Voice

    Real-time conversational telephone agent

    A dedicated voice agent answers calls on your office number, conducts natural inquiries, and schedules consultation slots. Call transcripts and structured protocols attach to the client record as soon as you accept the inquiry.

  • Isolation

    Multi-tenant database isolation as an invariant

    Every database query is strictly pinned to firm and matter ID; write access in matter chats is clamped to that individual case. A comprehensive invariant test suite proves scope pinning, role gating, and cross-firm isolation against live test databases rather than shallow mocks.

  • Compliance

    GDPR, EU operation, placeholders

    Workspace and data storage in the EU. In Juna chats, drafts and the deadline scan, a rule-based layer replaces names, email, phone, IBAN and other details with placeholders before text reaches a third-party language model; the approval card shows you the originals. Telephony, speech recognition and the search index over your documents need the original text. Files uploaded in chat expire automatically.

Architecture blueprints, security documentation, and audit findings are shared openly during our technical walk-through, including edge cases and ongoing hardening.

No Obligation

Request a Demo

Get in touch. We respond within one business day.

We use your details to handle your request. More in our privacy policy.

  • Free of charge and without obligation
  • Response within one business day
  • No disclosure to third parties