For this hackathon I have been rebuilding the project around one idea: a legal assistant should never say something it cannot back up. The earlier version could draft a case strategy from an FIR, but I could not honestly promise that every citation was real. So the new version works differently. The model no longer writes free text. It produces small claims, and each claim is checked against its source before anything reaches the advocate. Facts must appear in the FIR, statute mappings come from a table instead of the model’s memory, and every precedent must exist in the database with the quoted passage actually present in the judgment. Anything that fails a check is flagged or left out, with the reason shown.
I also added an audit mode. You paste any draft, even one written by another AI, and it tells you which citations hold up.
The models now run on Nebius Token Factory with NVIDIA Nemotron, and Tavily is used to look for later treatment of a case. I will share evaluation results once my final run finishes, because I only want to post numbers I can reproduce. If you work with Indian criminal law, I would really value your feedback.
Log in or sign up for Devpost to join the conversation.