Lawtionary (AMT)

The review problem

Lawtionary is a team prototype for Indonesian legal-draft review and source-backed terminology lookup. The application is designed as a local-first macOS workflow where a reviewer can inspect the evidence behind a suggestion.

My contribution

I worked directly on the application logic, legal-language data collection and preparation, AI/ML components, and local model integration. I also contributed to the discussions that shaped how the AI component should fit into the document-review workflow.

Keeping suggestions grounded

The application keeps two concerns distinct: dictionary lookup and contextual suggestion. Candidate retrieval is followed by structured outputs and evidence validation, with human review controls remaining in the flow. The prototype uses Swift, SwiftUI, AppKit, retrieval-augmented generation patterns, and local MLX/Qwen integration.

Prototype boundary

The result is a local macOS prototype that connects document review, terminology data, candidate retrieval, and local model output. It is a research and product prototype, not an autonomous legal advisor. It does not replace a lawyer’s judgment, and unsupported terms or suggestions should remain reviewable rather than being invented or silently rewritten.

What I learned

The project taught me to treat evidence and user control as part of the product architecture. A model output is only useful when the surrounding system makes its origin, confidence, and limits understandable.