Liveclassplatformsthatholdatscale.
WebRTC and SFU architecture, adaptive bitrate, pre-warmed regional infrastructure for scheduled classes, and interactive features decoupled from the media layer. Built for thousands of concurrent students per session.
Daily active users
Peak requests per minute
Products shipped
Downtime through migrations
Platforms at three inflection points.
Building a live class MVP
- Who
- Founders launching a first live class platform.
- Problem
- Need to ship a working live class experience without building global media infrastructure from scratch.
- What we do
- Application layer on top of Agora or AWS IVS. Class scheduling, attendance, basic interactive features. Live in 12-18 weeks.
Scaling past vendor limits
- Who
- Platforms that have outgrown the economics of pure vendor SDK usage.
- Problem
- Vendor billing has become the second-largest infrastructure line item. Time to control more of the stack.
- What we do
- Hybrid architecture. Vendor for the most demanding workloads, custom WebRTC + SFU for the rest. Significant cost reduction at scale.
Custom live class experiences
- Who
- Platforms with unique pedagogical features that vendor SDKs cannot support.
- Problem
- The needed UX requires deep control of the media stack, proctoring overlays, multi-camera teacher views, low-latency annotation.
- What we do
- Custom media stack with WebRTC + SFU + interactive layer designed for the specific pedagogy.
Where platforms break. And how we rebuild them.
Quality drops on slow connections
The pain: Students on weaker network drop in and out of the class. Teachers cannot see who is connected.
Our approach: Adaptive bitrate streaming that downgrades video quality before audio quality. Connection state monitoring surfaced to the teacher dashboard. Automatic reconnection with state preservation.
Pre-class join storms
The pain: Class starts at 4:00 PM. At 3:58 PM, 5,000 students try to join simultaneously. The signaling server saturates and half the joins fail.
Our approach: Pre-warm the SFU pool ahead of scheduled classes. Tiered signaling architecture that handles burst joins. Connection retry logic with exponential backoff client-side.
Recording and replay performance
The pain: Recordings take hours to be ready for replay. Students cannot review yesterday's class today.
Our approach: Server-side recording during the live session. Async transcoding pipeline that produces adaptive bitrate variants in parallel. Recording usually replay-ready within 15-30 minutes of class end.
Bandwidth costs spiraling
The pain: Egress bandwidth is the single largest infrastructure cost. Each new region multiplies the bill.
Our approach: Multi-CDN architecture for recorded content. Edge caching of common static assets. Live media routing through regional SFU clusters to minimize cross-region egress. Most platforms see 30-50% bandwidth cost reduction with the right architecture.
Methodology tuned for platforms at scale.
- 01
Media stack selection (week 1-2)
Pick the right media foundation for your scale and pedagogy. Vendor SDK, hybrid, or custom WebRTC. Decision is based on expected concurrency, interaction model, and unit economics, not on what is fashionable.
- 02
Application layer build (weeks 3-12)
Class scheduling, attendance tracking, interactive features (chat, polls, breakouts), teacher dashboard, student experience. The application layer is decoupled from the media layer so each can scale independently.
- 03
Recording and replay pipeline (weeks 8-16)
Server-side recording at the SFU layer. Async transcoding to adaptive bitrate variants. CDN distribution. Replay UX that feels native to the live class experience.
- 04
Scale hardening (weeks 12-20)
Pre-warm logic for scheduled classes. Multi-region SFU deployment. Monitoring tuned for media quality metrics (jitter, packet loss, glass-to-glass latency). Pre-event load testing on every major class.
250,000+ daily users. Multi-tenant by design.
Our multi-tenant EdTech platform powers white-label brands including Your CA Buddy and Youth Pathshala. It holds 250,000+ daily active users, 10 million requests per minute at peak, and has sustained zero downtime through three major scaling migrations. Every pattern on this page, the architecture, the decisions, the approach, has been battle-tested there first.
READ THE PLATFORM STORYHow the platform scaled from 20K to 250K daily active users over 3 years.Read case study →Questions founders ask about this.
How many concurrent students can a live class actually support?+
Depends on architecture. Peer-to-peer WebRTC: 8-12 participants. SFU (selective forwarding unit): hundreds per room with bidirectional audio/video. Hybrid SFU + broadcast: thousands of viewers per teacher with unidirectional video and lightweight interaction. We typically use the third pattern for K-12 and large-format classes.
Should we build on Agora, Twilio, AWS IVS, or roll our own?+
Most platforms should not roll their own at the start. Agora and AWS IVS handle the heavy media infrastructure. Build the application layer on top, class scheduling, attendance, recording, interactive features. Migrate to custom WebRTC only when usage patterns make a vendor uneconomical (typically past 50,000 hours of monthly streaming).
How do you handle interactive features like polls, chat, and breakouts?+
Separate the interactive layer from the media layer. Real-time chat over WebSockets. Polls and quizzes through a dedicated event service. Breakout rooms as nested SFU instances. The media SDK should not be doing interaction, keep them decoupled so they scale independently.
What is the right way to handle recording and replay?+
Server-side recording at the SFU layer, not client-side. Recording goes to object storage with a CDN in front. Generate adaptive bitrate variants asynchronously. Replay treats recordings as standard CDN-served video, no live infrastructure involved.
Latency targets for live class quality?+
End-to-end latency under 400ms for interactive classes (teacher needs to see student responses). Under 2-3 seconds is acceptable for broadcast-style classes with limited interaction. Above 5 seconds breaks engagement, students lose connection to the moment.