AI integration · Custom-built
Safe, affordable AI for a large music listings site
A large music site needed to process huge amounts of show and venue information, far too much to send through AI at a sensible cost. Ordinary code now handles more than 90% of it, and an isolated, metered AI agent handles only the parts that truly need it.
The problem
The site collects show and venue information on a very large scale. Running all of it through AI would have been the easy answer, but at that volume the cost would have been prohibitive.
The data is also public, written by anyone and fetched from anywhere. That makes prompt injection a real danger: text hidden in a page could try to give the AI instructions of its own, and an AI with access to the system could be talked into damaging it.
What we did
- Ordinary code first. We wrote traditional processing with proven tools such as regular expressions and page (DOM) parsers. It handles more than 90% of the data, quickly and at almost no cost.
- AI only where it is needed. The rest is show and time information written freely, with no fixed format. An AI agent reads it and pulls out the key fields, and that is all it is used for.
- Clean input. Everything sent to the agent is filtered and formatted first, which lowers the risk of prompt injection.
- A sealed agent. The agent runs in isolation, with no access to any system resource other than the input it is given.
- Checked output. Everything the agent returns is filtered again, so its results cannot harm the system.
- A meter on spending. The agent is metered, so a fault or an attack cannot run up unexpected charges.
What changed
The site processes its full volume of show and venue data at a fraction of the cost of an all-AI approach. The AI does the one job that ordinary code cannot, and even if it were fed malicious text, it has no way to reach or damage anything else.