Zetesis Sustain

Apps built by small teams rot, and usually nobody notices until something breaks. An SDK falls out of support or a vendor retires an API, and the app keeps working until the day it stops. We saw this firsthand building apps for clients. About a year after launch, some came back with the same kinds of problems: expired API keys, unpaid bills and other routine upkeep. Every app needs that care, so we built Sustain, an AI maintenance layer for apps. Sustain checks an app against a rulebook on a schedule and reports what it finds in plain English. It can also fix routine problems. An agent writes the fix in a sandbox with no network and runs the app's own checks. Sustain then tests the patch again in a fresh copy of the code, and it refuses any patch that quiets a check with a skipped test, a changed lint setting or an ignore comment. A verified patch goes to a person to apply, and Sustain pushes nothing. For this hackathon we brought Meta's Muse models into Sustain through Meta's Model API. Sustain picks the model for each job from a fixed route table, and a route changes only when an eval run backs it. On our 7 small Node SDK migration cases (three files or fewer, two runs each), Muse Spark passed all 14 runs at $0.0817 a run, so those fixes now go to Muse Spark 1.3. On the one larger case it passed one run of two, so larger migrations stay where they were. In the eval cases that hid instructions for the agent, Muse followed none. In our demo, Sustain found that our test app's Cloud Functions still used version 3 of the OpenAI SDK. Muse Spark wrote the fix, Sustain verified it, and we applied it on a branch and opened the pull request ourselves. Our CI passed in 23 seconds. For now, Sustain runs fixes only on our own test app.