PreFlight

PreFlight tells a creator which of two Reel openings will do better before posting, measured against that account's usual. Drop in two videos. Muse Spark watches the first five seconds of each and writes down, second by second, what is seen, written on screen, said and heard. Muse Glimmer 30B, fine-tuned with QLoRA for exactly this question and served on Modal, reads both write-ups and makes the call, with the record of calls as sure as it on accounts no model trained on. Muse Spark then explains why, and what could lift the other opening. Built during the hackathon from official data only. Instagram's Business Discovery API gave 62,482 Reels from 1,787 accounts with their view counts, and each Reel is labelled against its own account's median. It was tested twice on accounts no model saw: 207 pairs, then 419 more drawn after everything else was fixed. Over all 626 pairs from 90 accounts, the fine-tuned Glimmer picks the Reel that did better 67% of the time, against 57% for Muse Spark zero-shot and 50% for Glimmer before fine-tuning. In its surest quarter of calls it is right 82% of the time, and its pick typically got three times the views of the other. The second test also overturned our first design (two judges, "too close to call" when they disagreed), and the demo changed to match. Live demo: https://natanr123--preflight-web-web-site.us-east.modal.run