Sharing some learnings from my building process as I approach my Product Hunt launch date.
Here are four of the most important learnings thus far. #buildinpublic
Evals are essential to building a good AI product. Tabster uses AI to read receipts and process the data to generate an interactive version. I created an eval system that runs agains all receipts ever uploaded, so I can see how the model improves over time based on the underlying model changes (i.e. Gemini 2.0 > Gemini 3.0 Flash), and prompt changes. This leads to an iterative process where the product just gets more and more accurate. Right now the eval is showing 100% pass rate with perfect processing across thousands of receipts.
Understand user expectations. And exceed them. If users expect to only be able to select their own items and pay in a bill splitting app, include functionality to also select on behalf of others and request. Always raise the category expectations of excellence. If you can pleasantly surprise the user with functionality and quality, that's half the battle.
Respect the user's mental energy. This is more of a UX learning, but I have found that users respond better when you give them one main button or next step on a page, instead of giving them several things to process at once.
Don't wait to get feedback. I've continuously requested feedback from power users, and did not wait until the product was perfect to get it out there in the hands of real people. While this may have been humbling at times, I believe it allowed me to direct my work more efficiently towards real user requests.
Tabster is the best app I’ve found for splitting bills with friends, saves so much time and is easy to use!