The latest articles on DEV Community by Salahaldain Abduljalil (@salahaldainabduljalil). https://dev.to/salahaldainabduljalil https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F2529927%2Fc2964800-7960-4c69-991a-7bc42fcd3d9b.jpeg https://dev.to/salahaldainabduljalil en Salahaldain Abduljalil Thu, 17 Sep 2026 14:17:32 +0000 https://dev.to/salahaldainabduljalil/you-build-the-voice-ai-i-provide-the-cpaas-that-scales-it-2a24 https://dev.to/salahaldainabduljalil/you-build-the-voice-ai-i-provide-the-cpaas-that-scales-it-2a24 <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuxbdeoqjcmmy2qoekzgz.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuxbdeoqjcmmy2qoekzgz.png" alt=" " width="799" height="434"></a>Most companies building Voice AI products spend months (and<br> tens of thousands of dollars) building the telephony<br> infrastructure just to connect a phone call to their AI<br> system.</p> <p>This diagram shows the exact architecture I use to skip that<br> — a production-tested bridge that connects real phone calls<br> to any Voice AI engine in days, not months.</p> <p>If you're building a Voice AI product and don't want to<br> reinvent the telephony layer, this is the piece that's<br> usually the bottleneck.</p> <h1> VoiceAI #StartupInfra #B2B #SIP #CPaaS </h1> Salahaldain Abduljalil Thu, 17 Sep 2026 13:41:09 +0000 https://dev.to/salahaldainabduljalil/how-i-handle-1642-sfu-connections-at-6-cpu-on-a-29-server-kamailio-rtpengine-kernel--i https://dev.to/salahaldainabduljalil/how-i-handle-1642-sfu-connections-at-6-cpu-on-a-29-server-kamailio-rtpengine-kernel--i <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1ah43bso9c50thnv12vc.jpg" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1ah43bso9c50thnv12vc.jpg" alt=" " width="799" height="393"></a>How I Handle 1,642 WebRTC Connections at 6% CPU on a $29 Server - Pure mediasoup Benchmark</p> <p>In May I stress-tested the WebRTC layer of mena-stream.com.</p> <p>The result: 1,642 concurrent mediasoup SFU connections + 2,057 WebSocket signaling connections at 6% CPU, on a single Hetzner (8 vCPU, 16GB - $29/mo).</p> <p>Important: This May benchmark was PURE mediasoup + uWebSockets only. No RTPEngine, no Kamailio at that time.</p> <p>How I got that number with pure mediasoup:</p> <p>Tuned mediasoup workers, optimized uWebSockets.js signaling, and focused on SFU forwarding efficiency. No kernel RTP forwarding was used in this specific May test just kernel tuning and proprietary optimizations.</p> <p>The Grafana screenshot below is from that pure mediasoup run.</p> <p>What's new after May:</p> <p>After this benchmark, I added the full telephony layer on top:</p> <p>Kamailio + RTPEngine (xt_RTPENGINE kernel module) SIP gateway<br> DTMF (RFC 2833) pass-through verified<br> SIP-to-Mediasoup bridge via PlainTransport, so real phone calls join WebRTC rooms<br> GStreamer for recording, FFmpeg for smart rendering<br> Automated regression suite - 10/10 passing</p> <p>This new SIP layer is functionally working but its formal load test is still pending. I will publish a new benchmark with Kamailio + RTPEngine + mediasoup combined soon.</p> <p>Next: Combined load test + AF_XDP kernel-bypass.</p> <p>I submitted this architecture as a deep-dive to ClueCon 2026.</p> backend infrastructure networking performance