Pavo AI

Most engineering teams ship their three best guesses a quarter & hope one lands. Dashverse tested fifty ideas and shipped the one that won. Here's how 👇. @dashversetech runs Shortfree, a micro-drama app with 20M+ viewers 📱. Autoplay is one of its most important surfaces. It moves activation and retention directly. They already had a sophisticated recommendation system behind it: a contextual bandit balancing exploration (try something new) against exploitation (serve proven hits), tuned over several versions. Further than most teams ever get. And like every good team, they had no shortage of ideas to make it better. The hard part is knowing which ones will actually win an A/B test - when you only get three slots a quarter ⏳. So Dashverse and @Pavo_AI built what most companies only talk about: an offline setup that simulates any policy on logged data, no A/B test required. And it worked ✅. 🔍 Pavo's system understanding read their warehouse and codebase and built a detailed picture of how the recommender actually behaves - metric definitions, past experiments, what the ranker really does in production. 🏭 Pavo's applied science factory took it from there: found the gaps, generated 50+ candidate policies, built and validated the offline eval, screened everything, found two that beat the baseline, sent the PR. 📈 Offline eval put the winning policy ~2.6pp over baseline (estimate on logged data) and the live A/B gave +3.1pp activation lift! And the system knowledge carries over - the same understanding of Dashverse's user behaviours, content and metrics now applies to notifications, search, and every other surface 🔁. Full case study in the comments. Thanks to @dementorSam, HARSHIT and the team for a great collaboration!