Kyaloon started with a question shoppers in India ask all the time: “kya loon?” What should I buy? We began with supplements, and tried to get an AI to judge them fairly. Each try failed, and each failure left us one rule.
One AI call reads the label, searches the web, and gives a score.
A good product could score badly just because its lab report was not online. Most Indian brands do not put theirs online.
Write a rulebook for each category, and let the AI grade against it.
One call did the research, the judging and the writing. When a score looked wrong, we could not tell you which part went wrong.
Skip the AI. Compute the grade in code.
A label written a little differently could flip the verdict. On one product, a note in brackets turned the best ingredient into the top concern.
Give the AI a fresh brief for each product.
To keep the brief short, we cut the research. A claim on one label went unchallenged, because the evidence against it had been cut.
Research every ingredient first, then grade in one step.
Run the same product twice, and it could be judged against a different set of health goals. A rule we had changed was still applied in one place, and nothing flagged it.
Every version failed for the same reason underneath: the rules lived inside the AI’s instructions, where nobody could check them. So we are building an evaluation system for any kind of product, where the rules live outside the AI and every answer is checked against them.
It cannot evaluate a product yet. The AI judge comes next. Open source: github.com/Ronitron007/product-eval-harness