A Harvard Law dropout has raised $6 million for Blue Voice, an app that provides real-time legal and policy guidance to police officers in the field. The pitch is Harvey AI for cops: a model that knows use-of-force law, departmental policy, and case precedent, available in the moment a decision must be made. SignalFire and Las Olas VC led the seed round. The startup is entering a domain where the consequences of a wrong answer are irreversible.

The LLM Evaluation Problem, Applied to a Gun

A 2026 paper on arXiv by Sachdeva and Boudol, 'Rating the Raters,' identifies a structural problem with using large language models as evaluators: the model sits on every side of the assessment simultaneously, as examinee, as judge, and increasingly as the entity being judged by other models. The feedback loops this creates are not well understood. Blue Voice is proposing to put that kind of system into one of the highest-stakes evaluation moments in civilian life: an officer deciding, in real time, whether an action is legally permissible. The model's confidence interval is not something a person can intuit in the field. A related arXiv paper on 'Information-Seeking Psychology for Human-Centered XAI' finds that people do not seek explanations equally. Under stress, they seek confirmation. An AI that tells an officer what they can do, in a moment when the officer is already committed to doing it, is not a check on behavior. It is a permission slip with a latency advantage.

Who the App Actually Serves

That is the structural tension Blue Voice has not yet answered publicly. Real-time guidance for officers could reduce wrongful stops, or it could provide institutional cover for them. The difference depends entirely on training data, on which legal interpretations the model weights as authoritative, and on who audits the outputs. . SignalFire has a strong record in enterprise software, but policing is not enterprise software. The accountability structures are different. The clients cannot fire the product.