Matt Lawler

Matt Lawler helped build Joey, AssemblyAI’s AI support agent, with a team seeking to make technical assistance scale beyond the capacity of its onboarding engineers. He led forward deployed engineering at AssemblyAI during the 2026 AI Engineer World’s Fair. His work connects speech AI product quality with the practical demands of helping developers adopt it.

In 2024, Lawler contributed to English quality assurance for Universal-1, AssemblyAI’s multilingual speech-recognition model. His later work on Joey addressed a different part of the developer experience: how to help customers use speech technology when demand outgrows the team available to support them.

Turning customer knowledge into software

AssemblyAI was receiving roughly 1,000 API signups a day with one onboarding engineer. An off-the-shelf support bot resolved about 10% of conversations. Lawler and his team built Joey so they could directly change its instructions, retrieval, tools, and deployment. In his account of building Joey, Lawler reported that the agent resolved 80% of conversations end to end at approximately $700 a month in operating costs.

Three choices explain his approach:

• Documentation-grounded support: Built with the Claude Agent SDK, Joey combines documentation checked out as local Markdown, embedding retrieval, agentic file search, and a large CLAUDE.md instruction file. These give the agent access to product knowledge and let the team change how it finds and uses that knowledge. Deployment on Railway allowed fixes to ship in about 30 seconds.

• Escalations as a roadmap: Cases that still require a human identify what the agent needs to handle next. Lawler uses those unresolved requests to guide improvements, giving customer support a direct role in the agent’s development. This expresses his argument that forward deployed engineers should automate recurring parts of their own work.

• Using the product customers build with: Lawler added a voice interface to Joey using AssemblyAI’s Voice Agent API, connecting speech recognition, a language model, and speech synthesis through one WebSocket. Building with the same technology customers use makes the engineering team users of its own product and connects support automation with hands-on product experience.