CLOCK Counter-Linguistic Operations for Confession & Knowledge
Inspiration Every 11 seconds, someone loses money to a scam. Not because they're gullible — because AI voice cloning and number spoofing have made scams indistinguishable from real calls. Your phone rings. It's your daughter's number. She's crying. She needs ₹50,000 now. Don't tell anyone.
Every anti-scam tool today does the same thing: it warns the victim. But a warning is useless when emotional hijacking has already shut down logical thinking. We asked a different question — what if instead of warning the victim, we fought the scammer? Scammers use AI. So we built an AI that interrogates them.
What it does CLOCK engages scam bots in a live forensic interrogation using four techniques: reverse Turing tests, meta-questions, contradiction cascades, and non-sequitur traps. It asks the bot what it ate for breakfast, to spell words backwards, to name three things on its desk. A human answers in three words. A bot deflects, over-explains, or contradicts itself.
When the bot breaks character, CLOCK captures it in a signed, SHA-256-verified AI Confession Report — timestamped, tamper-evident, exportable as PDF, JSON, or TXT. Evidence for regulators, journalists, and law enforcement. The output isn't a warning. It's testimony.
How we built it We used Groq's Llama 3.3 70B for sub-second interrogation reasoning — critical, because counter-manipulation is a real-time problem. The scam bot and interrogator run in a dueling loop, with a multi-signal confession detector scoring identity breaks, contradictions, repetitions, refusals, and latency tells. The dashboard is built in Streamlit, with a live Threat Level Index that visualizes behavioral scam patterns (authority, urgency, isolation, extraction, fear) as the call unfolds. Reports are generated with ReportLab and cryptographically hashed for tamper-evidence.
Challenges we ran into The hardest part wasn't the code — it was the prompts. Our first interrogator was too polite; the scam bot hung up in three turns. We iterated over twenty times, tuning tone, chaining logic, and adding defenses to the scam bot so it fought back realistically. We also had to make confession detection strict enough to avoid false positives — a single signal never triggers a claim.
Accomplishments that we're proud of We built a working, end-to-end product in four days: live interrogation, real-time threat scoring, and cryptographically signed evidence export. The scam bot fights back for 8–10 turns before breaking — which makes the confession feel earned, not scripted.
What we learned AI safety isn't just about defending against AI — sometimes it's about using AI to hold other AI accountable. We also learned that in real-time systems, prompt quality matters more than model size. Groq's speed made the entire concept viable.
What's next for CLOCK Voice integration with Whisper for live phone calls. Carrier and bank SDK partnerships. A public evidence ledger for journalists and regulators. And eventually — autonomous honeypots that tie up scam operations at scale, turning the scammers' own tools against them.
We don't just warn victims. We make the AI testify against itself.
Log in or sign up for Devpost to join the conversation.