Brivva Tech

2026

Real-time multilingual live-commerce broadcasting. One stream, multiple markets, host's voice preserved.

Role
Co-founder + tech lead
Stack
Rust Cloudflare Workers AWS Fargate Terraform
GitHub →

Problem

Korean live-commerce hosts want to reach Chinese, Japanese, and Southeast Asian audiences — but translation options are either post-processed (loses the live moment) or captions-only (viewers scroll past untranslated hosts). Existing platforms like Prism handle broadcasting; Dubly handles dubbing. Nothing does both live, at the same time, with the host’s own cloned voice.

Approach

  • Host streams once in source language (Korean/English)
  • Server-rs pipeline (Rust + Axum + tokio) runs STT → translation → TTS with the host’s cloned voice, per target language
  • FFmpeg per-target encodes with burned-in captions for each destination
  • Multi-platform simultaneous RTMP fan-out: Grip (Korean commerce), TikTok, YouTube — one broadcast, N platforms
  • Cloudflare Workers + D1 handle sessions, billing metrics, OAuth
  • Soniox real-time STT with native translation; ElevenLabs Flash v2.5 for low-latency TTS

Outcome

  • Launched May 2026 with first B2B broadcaster
  • Pipeline targets for scale: 10 clients × 3 languages = 60 output-minutes per source-minute, ~90% margin at $1.50/output-min
  • 8 vCPU / 16 GB Fargate task handles En → Ko+Zh+Ja + passthrough on Grip + TikTok + YouTube concurrently
  • Clean 4-layer architecture (core → shared → features → orchestration) + three-tier testing discipline (unit + integration + e2e)

What I learned

  • Real-time voice translation is four chained failure points (STT / translate / TTS / RTMP) — observability at every stage is non-negotiable
  • Vendor-specific encoder quirks matter: AWS IVS rejects native ffmpeg RTMP but accepts librtmp; FFmpeg 7.1+ is required for the drawtext-with-harfbuzz chain
  • Vertical focus (Korean live commerce) beats horizontal sprawl (generic “translation for everyone”) — specific platform integrations are the moat, not the ML model

Status

Development stopped in May 2026 when the funding round did not close. The code is open source: github.com/TyposBro/brivva — the Rust media pipeline, the Cloudflare Workers API, the shared contract package, and the OpenTofu infrastructure.

A recorded end-to-end test from April 2026 is on YouTube: a real host, pushed through the pipeline, with the translated output going out as a live stream. Test footage, not a product demo.