LDA: Private Legal AI Workspace

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for LDA: Private Legal AI Workspace

Inspiration

Lawyers want the productivity benefits of modern AI, but legal documents carry client names, deal terms, personal data, bank details, signatures, and other sensitive information. Uploading an untouched agreement to a general-purpose cloud service can create confidentiality and data-protection risk. LDA creates a practical privacy boundary on the lawyer's own Mac.

What it does

LDA imports legal documents, detects sensitive entities locally, and gives the lawyer a review screen before replacing approved items with typed placeholders such as {PERSON_1} and {COMPANY_1}. The anonymized document can then be used in an external AI workflow. When the work returns, LDA restores the original values locally from an encrypted mapping.

The app also maintains encrypted matter workspaces, supports rename and archive workflows, recovers interrupted sessions, and can fill draft documents from encrypted client profiles. It supports text, DOCX, and PDF workflows and runs a local GGUF model through llama.cpp and Metal. Its macOS App Sandbox has no network entitlement, so the operating system blocks outbound network access.

How we built it

The native app is written in Swift and SwiftUI. LDACore combines deterministic detectors for structured data with on-device language-model inference for contextual entities. Mapping and matter data are stored in versioned AES-GCM containers, with app keys protected by the macOS Keychain. Document processing supports reversible tokenization, review-first restoration, PDF redaction, and form filling. The release is signed with Developer ID, notarized by Apple, and validated by Gatekeeper.

During OpenAI Build Week, Codex helped turn an existing technical app into a coherent guided product. The July 20 extension added the matter workspace, rename and archive actions, session recovery, privacy-safe metadata migrations, workflow routing, extensive regression tests, and hardened packaging. Codex accelerated codebase tracing, implementation, test design, code review, and the notarized release workflow. The entrant made the key product and privacy decisions and verified the resulting behavior.

GPT-5.6 was the selected Codex model for the Build Week extension.

Challenges

The hardest problem was preserving a simple experience while maintaining a strict local privacy boundary. Matter names and lifecycle state had to remain usable without leaking client data into ordinary preferences or logs. Rename, archive, restore, and deletion also had to stay consistent across encrypted mappings and resumable sessions. Packaging a 2.5 GB on-device model from an iCloud-synced checkout introduced additional signing and filesystem edge cases.

Accomplishments

• A guided workflow that makes a complex privacy process understandable.
• Reversible anonymization with local review and encrypted mappings.
• Matter rename, archive, unarchive, and interruption recovery.
• Fully offline App Sandbox entitlements with no network capability.
• A signed and Apple-notarized macOS application.
• 808 passing automated tests across core, UI logic, security stores, CLI, MCP, document formats, packaging, and entitlements.

What we learned

Privacy is not one feature. It shapes storage, UI labels, logs, recovery, testing, and distribution. Guided workflows are especially important in legal software because a technically correct tool still fails if the user cannot see what will happen before applying a change. We also learned that local AI product quality depends as much on packaging and recovery as it does on inference.

What's next

Next steps include a smaller downloadable model, broader language evaluation, team-safe profile exchange, more PDF form controls, signed automatic updates, and structured audit exports that reveal process metadata without exposing client content.

Built With

• app
• codex
• cryptokit
• gpt-5.6
• llama.cpp
• macos
• metal
• pdfkit
• swift
• swiftui