Websitesthat actuallyship.

The problem

No single message in a real scam looks damning. A KYC text, a bank claim, a link, a countdown, then a request for a code: each step alone is weak evidence, so detectors that judge one message at a time miss it. The order the steps arrive in is the tell.

What we built

• Trust Drift: a live risk reading that changes as the conversation goes on and shows what moved it, rather than a fixed scam score.
• An attack-progression model across seven stages, from first contact to extraction, where each new stage counts for more when earlier ones have already happened.
• Scam DNA that matches behaviour rather than wording, so a new script with the same moves still matches, plus a trust graph of who claims what and asks for what.
• A family guardian without surveillance: nobody watches by default, and when risk turns serious it asks first, then shares reasons, not the conversation.
• Detection runs on the device with explainable rules, with an optional Claude read of new messages.

The problem

In Indian classical singing, much of the meaning lives in the movement between notes, the gamakas. Standard notation and ordinary pitch tools throw that movement away.

What we built

• Pitch measured in cents above the singer's tonic, with every 12 ms of audio drawn as a point in 3D space.
• Automatic tonic detection, which found the same tonic, 149.3 Hz, independently on two different recordings.
• Optional source separation with Demucs, so songs with instruments are analysed on the voice alone.
• Drag-and-drop analysis of your own recordings in WAV, MP3, OGG, FLAC or AIFF.
• Uploads that go straight to cloud storage, which gets full-length songs past the 4.5 MB serverless limit.