Hey Hunters Simon here, solo dev on Oracleodds AI. I spent a year building a sports prediction engine and then hit the problem every AI product in this category has: there's no way for a user to tell a real model from a confident-sounding one. The entire industry runs on screenshotted wins and deleted losses. So this version is built around the opposite instinct. Every pick opens up to show its reasoning recent form, attack strength against that specific defence, the league's scoring baseline, rest days, and how the market has moved since we made the call. Behind that is a public performance board broken out by confidence band, month, sport and league, plus a head-to-head archive of 13,700+ finished fixtures so you can check a matchup's history without leaving. The part I'd rather not write: our confidence scores run hot at every band when we say 70%, we don't hit 70%. And we recently found a settlement bug that had been inflating our own published accuracy. We fixed it, re-graded 703 predictions, and republished the corrected numbers rather than quietly moving on. I'd genuinely rather lose signups to that than win them on a number I can't defend. Happy to get into the modelling stack (Poisson/Dixon-Coles/ELO ensemble, ~18 services, self-learning settlement loop) or the calibration problem specifically it's the thing I'm most stuck on. If you've shipped a probabilistic product: how do you present confidence to users without implying more precision than you have?

OracleOdds AI

The AI sports model that grades itself in public

The AI sports model that grades itself in public