SENTINEL — Adversarial AI Court for Bodycam

Tens of millions of police interactions are recorded on bodycams every year across the US, EU, and Italy. Less than 1% are ever audited. When a complaint arrives months later, the officer's written report becomes the de-facto ground truth - and lawsuits, careers, and lives are decided on one-sided narratives. SENTINEL is the audit layer that should have always existed. It is a three-layer adversarial AI court that processes every recording end-to-end: LAYER 1 - RAPID (sub-2 seconds per utterance). Speechmatics streams diarized transcription with sentiment, translation, and topic detection. A Router agent on Featherless triages each utterance against jurisdictional case law via FAISS retrieval, then a Rapid Prosecution agent emits a live alert in the UI. Alerts flag both officer violations and citizen violations - the system is bidirectional. LAYER 2 - DEEP COUNCIL (every 10 utterances or session end). A full Prosecution agent on Featherless gpt-oss-120b argues for the citizen. A Defense agent on Featherless gemma-3-26B argues for the officer. A Visual Context agent on Gemini 3.1 Pro analyzes the bodycam video directly. LAYER 3 - VERDICT (Gemini 3.1 Pro). The Judge weighs both sides for each alleged rule, issues per-rule rulings with prosecution_weight and defense_weight, and produces a defensible headline. The Judge returns officer_justified in demo cases - including one where drugs found in the vehicle established probable cause for a search the Prosecution had flagged as a 4th Amendment violation. It is accountability for both sides. Three jurisdictions ship today (US, EU, Italy). Adding a new jurisdiction is one rule-pack JSON. The same engine powers a corporate_security vertical for EU GDPR / labor compliance - proof the architecture generalizes to healthcare, insurance, and HR workflows.