A second listener.
A moment to pause.
Scam callers use fear, urgency and secrecy. Raksha helps you recognise the pressure, find the words to end the call and bring in someone you trust. Try the guided demo, or connect your own Moss project to test live semantic retrieval.
Caller-ID tells you who is calling. Raksha helps you recognise the pressure.
A caller claims to be the police. A bank account will be frozen. A family member needs money now. Under pressure, even a careful person can skip the moment of verification.
Raksha looks for familiar combinations: authority → fear → secrecy → the ask. It shows the conversation evidence behind a warning and gives the person a concrete next step.
Fast recognition. Clear advice. A person you trust.
Speech becomes text in the browser (its built-in speech engine; Whisper for uploaded recordings). Live mode sends text to Raksha. Uploaded audio passes through the server to Groq for transcription. Call details stay in RAM for up to ten minutes after the call ends. Every fragment, even mid-sentence, is a query.
With Moss connected, each fragment is embedded and matched against the 409-line playbook inside the server process. A deterministic risk engine combines the evidence. Without a Moss connection, a clearly labeled text detector keeps the demo usable.
When pressure meets an ask, the shield interrupts: it says stop, gives you one sentence to end the call, and alerts your guardian. An optional AI coach adds context outside the detection path.
A scam is won or lost inside one sentence.
“Read me the six digits, quickly.” A warning is useful while someone can still pause. Moss keeps embedding and search inside the server process, so detection does not wait for a remote vector database or language-model verdict. The Lab separates retrieval, analysis and network timings. Speech services and network conditions still affect the experience.
Historical Moss timing comes from the committed 264-utterance evaluation, not the current demo session. Other timings are illustrative. Run the latency lab to measure this server and your network separately.
Four ways the retrieval layer does real work here.
The curated scam playbook lives in a Moss cloud index, loaded into your session after connection and queried in-memory with a raw cosine score for every fragment.
Substantive caller turns are written into a local, in-process Moss session tagged with the call id. The guardian can ask “what did they ask for?” and get filtered semantic recall. This memory is not published to Moss Cloud; its turns are removed at call end.
The shared-runtime integration supports explicit caller reports to a second Moss index and automatic refresh. This depends on shared-project credits. Publishing is disabled for personal project sessions.
The shared runtime can combine playbook and community indexes in one query. Personal sessions query only their verified playbook. Benign look-alikes help the risk engine distinguish legitimate conversations.
The circle of trust
Scams work by isolating the victim: “tell no one”. Raksha breaks the isolation. A family code links a parent’s phone to a son or daughter’s dashboard. When risk rises they see the transcript highlights, can speak through the shield, and can ask the call’s memory a question. Human and agent, in the same session.
Open the guardian viewBuilt to learn from real use
- Today: a browser prototype with guided calls, recording uploads, live speech input and a guardian view.
- Next: consented testing with families, more accents and languages, and stronger guardian invitations.
- Before broader use: evaluate real-world false alarms, strengthen abuse controls and add shared state before scaling.
Raksha is a decision aid. It does not intercept cellular calls, verify identities or guarantee fraud prevention.