Most sleep advice is based on population averages: dim the lights, keep a routine, avoid stimulation. Sleep Tight asks a more personal question: what if your bedtime routine could learn from your actual day, your actual watch signals, and your own repeated sleep outcomes?
Sleep Tight is a local-first bedtime personalization tool for people who want a calmer, more targeted wind-down routine that adapts to their lifestyle without sending intimate health data to a remote service.
Each night, the Android phone exports the last 24 hours of available Health Connect vitals to a MacBook. At the same time, the Galaxy Watch records a bounded 10-minute pre-sleep heart-rate window. At 10:05 p.m., Codex Scheduled reads the latest local data, analyzes sleep quality where records are available, and chooses a safe background-sound and lamp-fade plan.
The room response is intentionally bounded: the lamp dims warmer and lower, background sound fades out, and the system avoids reacting to uncertain consumer sleep-stage estimates during sleep. Over multiple nights, Sleep Tight generates monthly learning logs that summarize what Codex learned and how the next routine should be personalized.
The repo includes the phone app, watch app, Mac receiver, local dashboards, lamp simulation, 30-day synthetic sleep report, future nudge simulator, setup scripts, tests, and a GitHub Pages walkthrough for judges.
Codex was the main build partner. It helped implement the Mac receiver, the phone/watch/Mac architecture, dashboards, lamp simulation, monthly report, setup scripts, tests, README, slides, and GitHub Pages walkthrough. GPT-5.6 played an integral role during research: it helped us reject unreliable real-time sleep-stage intervention and focus the product on the safer, more useful moment right before sleep.
Log in or sign up for Devpost to join the conversation.