A cloned voice can sound like your son, your boss or your best friend. KinCheck never judges the voice: it asks the real person.
Live call. A Vapi agent answers and transcribes with name-boosted Deepgram nova-3. Superlinked's serverless inference (GLiNER) catches the identity claim, money, urgency and secrecy in ~150 ms. Cognee's knowledge graph turns "your teammate from the hackathon" into a real contact. KinCheck asks that person "Is this you?" on their own channels at once: WhatsApp via Beeper and Telegram.
One decision. SingleStore holds every call: one atomic update lets exactly one answer win. No ends the call, Yes forwards it, and every step is audited. SingleStore vector search over Superlinked bge-m3 embeddings names known scam scripts.
Trust and setup. Persona checks identity at sign-up (hosted flow, sandbox). The app manages trusted contacts, the Cognee memory graph and expected calls from a mailbox scan (sample): a real courier passes, a fake fee is blocked. A consented cloned-voice call is proven end to end.
Measured. On our 34-caller drill set, rules flagged 16 of 18 impersonations with 1 of 12 false alarms; after the fix, 18 of 18 and 0 of 12. A real Firedrill run of 3 hard callers held both scams and wrongly challenged one genuine caller: the false alarm we then fixed. Amplitude Agent Analytics records each call with four rule evals: never offered a name, never agreed to pay, only verified actions, asked the real person. The model talks; only code acts.