About the project
Problem. Every creator gets the same advice: post reels, hook in the first second, post at 6pm. It is an average over millions of accounts, so it is wrong for almost every specific one. Instagram analytics show what happened to each post, not which repeatable choices moved the numbers.
Who it helps. Independent creators with a public account and 100 posts of history, and the agencies and brand teams that manage them. No login, no API approval.
Solution. Hindsight scrapes an account's last 100 public posts through Apify, turns each post into 25 computed features plus four LLM labels (hook, tone, call to action, topic), normalizes the outcome per account, excludes collab and pinned posts, and bootstraps every feature value 1,000 times for an effect with a 95% confidence interval. Effects that do not reach significance are labeled inconclusive with a reason. One synthesis call writes the headline and five post briefs from the statistics table alone; the model never sees raw posts, and every percentage is overwritten from the table in code.
Impact. Same feature, opposite sign: educational tone earns NASA 62% more engagement and costs a 1.3M-follower tech creator 44% of his views. In under two minutes a creator learns which rules are real for their audience and which are noise, with receipts.