George Guimarães

AI systems are usually complex: Python APIs, model servers, queues, multiple frontends, and way too many moving parts.

At work, we took a different path. We built a full-stack AI system in Elixir that serves over 3 million customers—handling everything from LLM-based chat to GPU-powered image moderation, all in one cohesive codebase.

In this talk, I’ll show how Elixir helped us: • Build a production-ready chat agent with multi-agent orchestration • Run neural networks on GPUs directly from Elixir using Bumblebee • Deliver real-time insights via LiveView dashboards • Avoid the overhead of microservices while scaling reliably

If you’re wondering whether Elixir can power serious AI at scale—this is living proof.

Key Takeaways:

• You can build and scale real AI systems in Elixir—LLMs, image models, dashboards, and all
• Elixir’s concurrency model and GPU integration (via Nx + Bumblebee) unlock surprising power
• LiveView enables full-featured, real-time operator UIs without extra frontend stacks
• A single, unified codebase can replace the complexity of multiple Python services
• Elixir is not just web/backend—it’s production-grade AI at scale

Target Audience:

• Elixir developers curious about AI/ML applications
• Engineers building or scaling AI-powered features (chatbots, moderation, etc.)
• Tech leads exploring alternatives to Python-heavy AI stacks
• Anyone interested in running LLMs, neural networks, and dashboards in one unified system
• Builders who care about simplicity, performance, and observability in production