Picture this: Your papa gets a WhatsApp message saying he won a lottery. He never bought a ticket, but the message looks super official.
Normally, heβd call you or a tech-savvy cousin to ask, "Beta, is this real?" But what if you are busy? What if he just clicks the link?
I wanted to fix this. What if he could just open an app and talk to it in natural Hindi, exactly like he talks to you?
That is exactly why I built Suraksha Saathi for the #VoiceForBharat 10 Days of AI Voice Agents Challenge. Here is how I went from an empty folder to a fully working anti-fraud AI in just 10 days.
β‘ The Engine Under the Hood
Before we get into the days, here is the simple loop of how it works:
You speak β Deepgram listens β Google Gemini thinks β Murf Falcon speaks back β LiveKit connects it all without lag.
A quick note on speed: If a voice agent takes 3 seconds to reply, it feels like a broken walkie-talkie. I used Murf Falconβthe fastest TTS API available right nowβand it genuinely made the AI feel like a real, instant phone call.
π
The 10-Day Journey
Days 1 & 2: Teaching it Hindi (Properly)
At first, the AI sounded like an American trying really hard to speak Hindi (reading "karu" instead of "ΰ€ΰ€°ΰ₯ΰ€"). The fix? I forced the AI to only output in the Devanagari script. Suddenly, it sounded fluent! I also gave it strict rules: Never ask for an OTP, and never promise a scheme is 100% safe.
Day 3: The Glowing Diya UI
Instead of using boring jumping audio bars on the screen, I built a custom animated Diya (oil lamp). It glows steadily when listening and flickers when speaking. Why a diya? Because its whole job is to bring light to the darkβwhich is exactly what this agent does for scams.
Days 4 & 5: Giving it a Memory
I connected a database so the AI remembers returning callers. But I added a consent rule: It actually asks, "Kya main aapka naam yaad rakhun?" (Can I remember your name?). If you say yes, it greets you by name the next time you call.
Day 7: The Panic Button
If a caller is actively losing money, the AI quietly pings a human fraud team on a private Discord channel in the background, all while keeping the caller calm on the phone.
Day 8: The Command Center
I built a live dashboard to track what kind of scams people are reporting. Now I can visually see if "Fake FedEx" or "Digital Arrest" scams are spiking in real-time.
Day 9: The Handoff Magic
If things get really serious (like a hacked bank account), a friendly guide isn't enough. I wrote code that swaps the voice engine mid-call. The friendly AI (Anisha) instantly transfers the call to a strict Fraud Investigator (Samar). The voice literally changes on the spot without the call dropping!
Day 10: Sharing it with you! (Writing this post).
π₯ The Midnight Bug That Almost Broke Me
On Day 4, I spent hours staring at a 400 Bad Request error at 11 PM.
I was trying to make the AI "look up" the user in the database before it even said hello. Gemini hated that. The fix? I stopped trying to make the AI do everything. I just wrote a simple Python script to look up the user first, and handed that info to the AI.
Lesson learned: Sometimes the fix isn't writing more AI prompts, it's just writing good old-fashioned code.
π Try it yourself
If you want to see the code or build your own, everything is public here:
π [Insert your GitHub Link: https://github.com/Bedi-Builds/murf-livekit-starter/commits/main/]
Ten days ago, this was just an empty starter template. Today, itβs an AI that speaks real Hindi, remembers the people who trust it, knows when to call a human, and could actually stop someone from losing their savings.
#MurfAI #MurfFalcon #LiveKit #Gemini #BuildInPublic
Top comments (0)