Leonardo Rodriguez

Mirada is a new answer to the oldest problem in cheap eye tracking: drift. Webcam gaze degrades within minutes because one-pose calibration memorizes a single head position — move a millimeter and the map is wrong. Our innovation, semantic recalibration, closes that loop: a local ~30fps gaze model corrected by sparse semantic anchors. Muse Spark reads screenshots plus interaction context and infers where you almost certainly looked; SAM 3.1 pinpoints it to the pixel; a tiny bias-only online corrector absorbs the fix — never an LLM per frame. Sessions replay offline under competing strategies (pilot, n=2: anchors cut mean error 406→397px while uncorrected drift explodes 63→941px in 15 min). The same anchors could correct headset gaze drift next. The innovation ships as a gaze communication board on an ordinary $30 webcam: look at a cell, close your eyes ~0.4s, and the computer speaks. Eight huge phrase cells with offline speech (EN/ES, editable), a full gaze keyboard with word completion and SEND, and an APPS page launching real apps (YouTube, browser, Discord…) where your gaze becomes the mouse — look freely, close to click — plus a floating dock to return anytime. The caregiver's mouse always works alongside. Built for people who depend on gaze — ALS, spinal injury, motor impairments — replacing $5,000 dedicated trackers with hardware they already own. Fully offline live loop, no API key needed: any judge can verify it with any webcam. Paper draft and replay harness ship in the repo.