Infrastructure software for teams running AI in production.
We build small, self-hosted tools that solve one problem properly. The first one is llm-guard — a gateway that detects runaway AI agent spend while it is still happening, attributes LLM cost to the key that caused it, and enforces hard budgets. It runs on the Python standard library alone.
Get a cost diagnosis → · See llm-guard →
Why we started here
In 2026 the cost of AI stopped being a rounding error and became a line item somebody has to own. OWASP's research into agent execution budgets catalogues 63 confirmed production budget-overrun incidents across 21 orchestration frameworks, and puts Fortune 500 leakage at roughly $400M in unbudgeted spend.
The pattern behind almost all of it is the same: an agent loop re-sends its entire conversation history on every step, so the input-to-output ratio climbs, and by the time a daily budget cap fires the money is already gone.
Most tooling tells you what you spent. llm-guard is built to catch the loop while it is running.
What we care about
Narrow over broad. One problem, solved well, beats a platform that does everything adequately. A gateway that meters cost should not also try to be your vector database.
Boring underneath. No third-party runtime dependencies, no telemetry, no data leaving your network. A tool that sits on the request path and holds your API keys is a target; the smallest possible surface is a feature, not an aesthetic.
Honest numbers. If a model has no price on file, the report says unknown rather than guessing. A cost figure that is plausibly wrong is worse than no figure at all, because someone will make a decision on it.
Two ways to work with us
If your AI bill is unpredictable, we will find out why. A fixed-price diagnosis, delivered as a written report: what is driving it, what to change first, and what each change is worth per month. No infrastructure to deploy, read-only access to usage data, and no prompts or completions ever leave your account. How it works and what it costs →
If you would rather run it yourself, llm-guard is MIT licensed and complete. Zero dependencies, standard library only, budgets and runaway-loop detection included. There is no crippled edition and no feature reserved for paying customers.
We are a small team in Hong Kong, and we would rather talk to the people running this in production than ship a newsletter. hello@lye-labs.com