kyma-icp-agent
Sorts new signups into real customers worth a person's time and everything else.
The problem #
A developer platform gets thousands of signups, and a founder cannot read them all. Most never become customers, a few are exactly who the product is for, and eyeballing a signup log does not tell them apart. The cost of guessing wrong is a week spent on the wrong accounts.
The approach #
A triage agent scores each account against an ideal customer profile, and a deterministic evaluation harness grades the agent without asking another model for its opinion. The harness was built to find mistakes in the profile itself, not only in the code.
Where it is now #
Built in one night at a hackathon on 2026-08-25 and run against real production account data. Not yet wired into the product as a live step.
What happens if this works #
Customer research becomes a test suite. When the profile of an ideal customer is written down as something a harness can check, it can be argued with, corrected, and rerun every week.
Evidence #
- 2,610 accounts triaged in the first run, 2026-08-25
- 15 of 20 cases passing on the full local evaluation, with no model used as a judge