Detectif AI builds models that flag synthetic, cloned, or replayed speech on live calls and recordings, match a caller against an enrolled voiceprint, and label the other side of a line as a person, an IVR, voicemail, or another voice agent. The company is opening that stack to any developer through a public API and an MCP server, with docs and a free key, for banks, telecom fraud teams, and builders of voice agents that need a signal before a payment or account change goes through. Founder Tarini Sai Padmanabhuni started Detectif AI after a cloned voice imitating a relative convinced her grandfather to pay a ransom. TechCrunch reported that the company already processes more than 100,000 calls a month for financial institutions in India, running deepfake detection and speaker verification on AI voice agents used for debt collection and loan-document follow-ups. On Podonos's benchmark of 4,524 clips scored against private labels, Detectif AI caught 97.3 percent of the synthetic audio and posted 94.47 percent overall accuracy, above the published scores for Hive and Reality Defender. The company reports 24 milliseconds per clip, the lowest latency among commercial systems that published timing on that board. Podonos computed the accuracy. The latency figure is vendor-reported.