Suraksha Saathi: Building an AI Voice Agent in 10 Days

Picture a scenario where an elderly parent receives a WhatsApp message claiming they won a lottery despite never buying a ticket. Normally, they would call a tech-savvy family member to ask if the message is real, but busy schedules or clicking a malicious link can lead to disaster. A developer set out to solve this exact problem by building an application where family members can speak in natural Hindi.

This led to the creation of Suraksha Saathi for the #VoiceForBharat 10 Days of AI Voice Agents Challenge, progressing from an empty folder to a fully operational anti-fraud AI within 10 days. The technical engine relies on Deepgram for listening, Google Gemini for reasoning, Murf Falcon for speech synthesis, and LiveKit for seamless real-time connection without lags.

Speed is critical for voice agents, as a three-second delay feels like a broken walkie-talkie. The developer utilized Murf Falcon, currently the fastest text-to-speech API available, ensuring the AI delivered the experience of an instant phone call. During days one and two, the focus shifted to teaching the AI proper Hindi by forcing Devanagari script outputs to eliminate unnatural foreign accents, alongside strict rules against asking for OTPs or promising 100% security.

On day three, a custom animated Diya oil lamp interface replaced boring audio bars, glowing steadily while listening and flickering when speaking to symbolize bringing light into dark scams. Days four and five introduced a database connection for memory, requiring user consent through the phrase "Kya main aapka naam yaad rakhun?" before addressing returning callers by name.