Saksi listens to face-to-face sales conversations between a loan officer and a customer, and whispers to the officer what they are legally required to say but haven't yet. Most voice agents talk to a person. Saksi listens to two people and stays silent until something is missing. It uses AssemblyAI Universal Streaming with speaker diarization to tell officer from customer in real time, checks each officer utterance against five mandatory disclosures, and produces an evidence-backed compliance report afterwards — every requirement linked to the exact utterance that satisfied it. Stack: Go, React PWA, Flutter, Postgres. Building solo.