Real-Time Fraud Call Detection
Budget: ₹12,500 – ₹37,500 INR
I need a complete solution that flags fraudulent phone calls the moment they happen. The core requirement is to identify fraudulent calls in real-time while simultaneously transcribing the conversation. I want transcripts both as the call unfolds and again once the call ends, so any post-processing or enhanced language models can refine the text.
A clean, responsive web-based dashboard must surface everything important at a glance: live fraud alerts, confidence scores, the running transcript, and a searchable archive of past calls with their final fraud verdicts. Think Twilio, WebRTC or any SIP-compatible stack on the telephony side, paired with a speech-to-text engine such as Google Speech-to-Text, Whisper, or a comparable on-prem model. The fraud classification layer can be rule-based, machine-learning, or hybrid—what matters is low latency, high precision and recall, and the ability to retrain easily as new scam patterns emerge.
Deliverables
• End-to-end codebase (telephony bridge, transcription, fraud detection, dashboard)
• Docker-based deployment scripts or equivalent containerisation for smooth setup
• Minimal UI/UX styling so the dashboard is production-ready
• Documentation covering architecture, model training pipeline, and environment variables
• A short video or live demo proving the system raises an alert during a test fraudulent call
Acceptance criteria: under 500 ms average detection latency after a suspicious phrase, ≥90 % precision/recall on the provided validation set, and transcripts with at least 95 % word accuracy.
If additional insight features—such as trend reporting—can be toggled on later, all the better, but the immediate goal is real-time protection with clear visibility through the web dashboard.
A clean, responsive web-based dashboard must surface everything important at a glance: live fraud alerts, confidence scores, the running transcript, and a searchable archive of past calls with their final fraud verdicts. Think Twilio, WebRTC or any SIP-compatible stack on the telephony side, paired with a speech-to-text engine such as Google Speech-to-Text, Whisper, or a comparable on-prem model. The fraud classification layer can be rule-based, machine-learning, or hybrid—what matters is low latency, high precision and recall, and the ability to retrain easily as new scam patterns emerge.
Deliverables
• End-to-end codebase (telephony bridge, transcription, fraud detection, dashboard)
• Docker-based deployment scripts or equivalent containerisation for smooth setup
• Minimal UI/UX styling so the dashboard is production-ready
• Documentation covering architecture, model training pipeline, and environment variables
• A short video or live demo proving the system raises an alert during a test fraudulent call
Acceptance criteria: under 500 ms average detection latency after a suspicious phrase, ≥90 % precision/recall on the provided validation set, and transcripts with at least 95 % word accuracy.
If additional insight features—such as trend reporting—can be toggled on later, all the better, but the immediate goal is real-time protection with clear visibility through the web dashboard.