Scam texts don't look like scams anymore. They copy your toll agency, your bank and even your carrier, and they win by rushing you. OnyxFlame's Second Look answers T-Mobile's Safe Signal challenge with calm protection that families can feel before, during and after a threat. Before a threat arrives, Nemotron-built lessons teach people to spot scams in texts, voicemails and emails. During one, every message is checked privately, real threats get a clear warning with honest confidence, and tapping a scam link brings up a calm pause with the safe way forward instead of the scammer's site. Afterwards, a recovery timeline guides next steps and checks back in. Families can see whether loved ones are okay without reading their messages, and a scam-only Nemotron assistant answers questions using trusted sources.
We built it with Next.js and NVIDIA Nemotron. Nemotron 3.5 Lightning triages each message, and Nemotron 3.5 Super plans an investigation and calls real tools: domain age, lookalike detection, sender signals and known scam patterns. Safety rules then double-check the verdict and catch messages that try to trick the AI. Along the way we handled unpredictable model output, a model retired mid-hackathon and prompt-injection attacks. We also rebuilt a cluttered first design into one calm phone experience, and we evaluated honestly using blind test sets.
We learned that smart friction beats loud alerts and that AI agents need deterministic guardrails. Next, we want to bring Second Look to real phones and live calls, connect real carrier signals, add Spanish, and let one customer's report protect everyone.
