WalkAround — Voice Inspection Agent

Every day, millions of truck drivers inspect their vehicles under federal requirements, often with a clipboard in the rain and gloves on. Items get skipped, defects are reduced to a few words, and there is little proof of what was actually checked. The same problem appears in restaurant openings, lab procedures, and equipment walkthroughs: paperwork fails exactly when hands and eyes are busiest. WalkAround turns any inspection checklist into a voice agent. The inspector simply talks: “oil’s good,” “left front steer tire is down to eighty,” or “skip the horn.” The agent announces each item, asks the follow-up that matters, and converts every verdict into a structured tool call. It can be interrupted, jump between sections, or revise an earlier item without losing its place because the checklist is modeled as a state machine, not just a chat. At the end, WalkAround files a structured report. Every defect includes the inspector’s exact words, a timestamp, and a work order. Safety-critical decisions such as “out of service” are preserved explicitly. The voice loop runs through a single AssemblyAI Voice Agent API connection, using Universal-3.5 Pro streaming STT with domain key terms, semantic turn detection, JSON-Schema tool calling, and native TTS. Tools and system prompts are swapped section-by-section over the WebSocket using progressive tool reveal, so the agent cannot skip required steps or file a report before every checklist item is resolved. WalkAround is not limited to trucks. Users can paste any written checklist, and AssemblyAI’s LLM Gateway compiles it into a voice-walkable template. Built-in templates demonstrate multiple domains, and a scripted walkthrough supports users without a microphone. Built with AssemblyAI Voice Agent API, LLM Gateway, Next.js 15, and Vercel. MIT licensed, with 28 automated tests and a live end-to-end API test.