Video streaming that scales with your AI.
Without the SaaS tax.

Break-even sits around 1M monthly participant-minutes for most use cases. Below that, SaaS is faster to launch and the per-minute economics work. Above it, custom builds amortize sharply — Nucleus runs 600M+ min/mo at fractions of a cent per minute, where a SaaS bill at $0.005/min would be $3M/month. The decision frame is in the Build vs Buy section above.

WebRTC: sub-200ms when the SFU is regional to your viewers. LL-HLS / CMAF: 1–3 seconds depending on segment length and CDN configuration. Standard HLS / DASH: 6–10 seconds. We pick per use case — interactive flows on WebRTC, large-audience live on LL-HLS, mobile / regulated on standard HLS. MoQ Transport (IETF draft) targets sub-300ms with CDN-scale fanout, which is the 2026–27 candidate for the hybrid path.

Yes — we run roughly 25% of engagements as parallel-build migrations. New mediasoup / LiveKit SFU comes up next to the existing SaaS, traffic is split by room class or feature flag, the SaaS stays as fallback while metrics are validated, then traffic shifts over once latency and QoE match or beat the baseline. Typical migration window is 6–10 weeks.

Yes. mediasoup SFU instances run in each major region (us-east, us-west, eu-west, ap-southeast, ap-east) with a routing layer that pairs participants by latency to the nearest healthy SFU. Cross-region SFU pipes handle multi-region rooms when needed. This is the architecture Nucleus runs to hit 600M+ minutes/month.

AI runs as a parallel inference pipeline so the primary stream isn't blocked by ASR latency. Whisper Large-v3, Deepgram Nova-3, or NVIDIA Parakeet for transcription; SeamlessM4T or DeepL Voice for translation; custom moderation classifiers for content policy. Captions / translations are delivered as a separate data track over WebRTC or as a synchronized side-channel for HLS — budget is under 800ms end-to-end for the AI output.

When architected correctly, yes. Compliance is enforced inside each layer: encryption at rest and in transit, role-based access for recordings, audit logs on every join / leave, data residency pinned to region (us-east only, eu-west only, etc.), retention windows configurable per room class, BAA signed for HIPAA flows, DPIA documentation for GDPR. We sign Data Processing Agreements before any engagement.

Yes — Worldcast Live runs 10K+ concurrent per event today on LL-HLS over CDN; the architecture extends linearly via CDN origin scaling. For interactive WebRTC flows, the ceiling is per SFU shard (1500–2000 participants) with horizontal sharding for multi-room or large-audience hybrid configurations (host on WebRTC, audience on LL-HLS).

10–12 weeks for a Startup-tier scope (single room class, up to ~500 concurrent, single region). 14–18 weeks for Growth (multi-room, AI overlay, multi-region SFU, up to ~10K concurrent). 18–24 weeks+ for Enterprise (multi-region clusters, full AI pipeline, full compliance overlay, custom moderation models). Discovery call to first running room is typically 3–4 weeks regardless of tier.

You do. Models, training data, infrastructure code, operator UI, and AI pipelines are all delivered to your repositories under your name. Fora Soft retains no claim on the IP. The benefit of custom development over SaaS is exactly this: the streams, the recordings, and the unit economics live on your balance sheet rather than the vendor's.

Three shapes: handover to your in-house team with runbooks and on-call training (most common at Enterprise tier); ongoing SRE / AI-tuning retainer (typical at Growth tier when in-house streaming expertise isn't on the roadmap yet); or fixed-scope quarterly improvement cycles (new room classes, new AI features, codec migrations). All three are scoped after the initial build, not bundled.