AVbase

It holds the room's gear, signal flow, and patching, plus the troubleshooting approach I use on shows. Anyone can use it, from someone who has never run AV to an SME who needs a refresher on an unfamiliar room. Voice works in both directions, so you can easily get answers from anywhere.

Rooms with frequent technician turnover spend their senior staff's time training techs of varying ability. AVbase makes orientation as simple as opening a link. New freelancers get the room's signal flow in plain language. New hires can ask the questions they don't want to ask the senior tech. Senior techs can pick up techniques outside their own habits. In a rental setting, the same idea works per rack: scan a code on the case and ask the gear what it's patched to.

The end state isn't a phone screen. It's the room itself: a dedicated mic you switch on, answers through the PA, so you always know when it's listening and when it isn't.

• It troubleshoots in the order a tech would: power first, then physical signal, then routing and processing.
• It refuses to route through a connection that doesn't exist, and tells you what to patch first.
• It pushes back on unsound requests and explains the risk, instead of complying.
• It knows where its documentation ends. Engineer-level settings get acknowledged, not invented.
• It explains each piece of gear the first time it comes up, then uses the name alone.
• It takes the day's show as typed text, a PDF, a Word file, or a photo of the printed sheet, and walks the setup step by step.

The knowledge base is composed at startup from a shared gear catalog and a per-room profile, so a venue with several rooms shares one library of manuals and each room gets its own setup.

The room's content started as real venue documentation and went through a sanitization workflow I wrote, with every brand, product, and client swapped for invented ones. That's what lets a real venue's rooms go in without exposing the venue or its clients.

The app is a small FastAPI service in front of Claude. No framework, no build step.

An eval suite scores the room by category rather than as one pass rate, so a regression in pushback behavior can't hide behind correct lookups. Baseline snapshots are taken before every model or prompt change. The most useful checks have been my own conversations with it, asking the questions I'd ask a tech I was training.