Self-Learning Legal Content Improvement System

My current workflow pulls fresh regulatory updates, case-law developments, and legislation changes, then lets an AI model draft a summary that a lawyer refines. I now need this refinement loop to teach the model automatically so each new piece requires less intervention.

Your main focus will be automated learning from expert edits: capturing every insertion, deletion, and comment, translating those differences into training signals, and fine-tuning the summarisation model so it steadily improves on content accuracy, relevance, and overall readability. I already handle security and explainability elsewhere in the stack, but the code you deliver must slot into that framework without exposing data outside our private repo.

To make this work you will:

• Build a diff-tracker that logs lawyer edits against the original AI draft, preserving context and timestamps.
• Convert those logs into structured training examples and schedule incremental fine-tuning or reinforcement updates.
• Expose an API endpoint that returns the next-best summary plus metadata so the reviewer can interrogate and, if necessary, override any sentence.
• Package the training pipeline (likely Python + Hugging Face/LLM-based, but I’m open to suggestions) with clear documentation and unit tests so it runs inside our existing architecture.

Acceptance criteria: after integration and one month of live use, the average word-level edit distance between AI draft and final lawyer copy should drop by at least 30% while maintaining legal accuracy (we will measure with our internal checklist).

If this sounds within your skill set, outline your proposed approach, the tools you would use, and a rough timeline for delivery.