We all treat closed captions as the baseline for accessibility, but for millions of users, they’re not enough. Sign languages possess unique grammar, emotion, and context that plain text simply can’t capture. While human interpreters are the gold standard, they don’t scale for the massive world of live streaming. So, how can we do better? This talk chronicles our journey to build a system that renders expressive, real-time sign language avatars directly into a live HTTP Adaptive Streaming (HAS) feed. We broke this enormous challenge down into a four stage AI pipeline: Audio-to-Text, Text-to-Gloss (the lexical components of sign language), Gloss-to-Pose, and finally Pose-to-Avatar. But the real engineering puzzle wasn’t just the AI models, it was where to run them. Every decision in a distributed system is a trade-off. This talk is a deep dive into the system architecture and those tough decisions. We’ll explore questions like: Should Audio-to-Text run in the cloud for maximum accuracy, or can the edge handle it to reduce latency? How can we use edge servers to handle Text-to-Gloss translation to support regional sign language dialects? What are the performance-versus-personalization trade-offs of offloading pose generation and avatar rendering to the client device? This isn’t an AI-magic talk. It’s a practical look at the system level challenges of integrating a complex, real-time AI workload into the video pipelines we work with every day. Attendees will leave with a framework for thinking about how to build and deploy novel, real-time accessibility features, balancing latency, scalability, and user experience. This talk was presented at Demuxed 2025 in London, a conference by and for engineers working in video. Every year we host a conference with lots of great new talks like this – learn more at https://demuxed.com