CleanQuest

THE PROBLEM Commercial cleaning employs 2.4 million individual cleaners in the U.S. CleanQuest uses AI to gamify validating that work: clean, time, technique. Crews see where better products and better training make them more efficient and productive. - Most crews get the bare minimum: minimal feedback, recognition, or proof of work. - Companies lose accounts over "they didn't clean" disputes they cannot settle. HOW IT WORKS — AI AT EVERY STEP 1. Check in: pick a language (English, Spanish, Polish), tap dropdowns — site, sector, area, spot. Zero typing. - Consent reads aloud: faces never identified; scores never used for discipline. - The before photo is captured in-app. No gallery uploads. - On-device MediaPipe (EfficientDet-Lite) rejects wrong-room and person-in-photo shots on the spot. 2. Run: a locked 90-second timer — "Quality first. Time second." - A neural host hypes the cleaner every 6–8 seconds, in their language. Encouragement only, never coaching. - 715 pre-rendered neural clips (Kokoro-82M + native ES/PL voices) ship in-app; host works offline. - One tap on COMPLETE CHALLENGE captures the after photo. 3. Get scored by Muse Spark Open (Meta open weights), run on hackathon partner credits: - Baseten ($100 credits + 2M TPM) serves the model; Fireworks AI ($100 credits) fails over, via a serverless proxy. Zero user keys. - Meta Model API credits ($150) funded Muse Spark 1.3 + Muse Wonder usage during the build. - The model grades every surface in the photo pair: clean, partly clean, missed. - It counts objects cleared vs. objects remaining; the app weights the verdict: - Clean 50% · Technique 25% · Time 25% (the timer supplies Time). - SAM 3.1 segmentation (own key) or Muse box_2d outlines each surface on their photos. - The system scores. Nobody can edit a score. - Settings proves it live: run log + telemetry test, verified on both providers (98% - 1,405 tokens; 96% - 1,695 tokens). - Optional own key unlocks Muse Spark 1.3 scoring and Muse Voice Transcribe. - Can't judge fairly? NOT SCORED + one plain reason (too dark, glare, camera moved) + a retake button. 4. Train and compete: - Daily Deck: tap-through training cards, AGREE / DISAGREE, day streaks. - Dual leaderboards: Shift Challenge + Daily Deck (Quality / Time / Style). - Personal bests, streaks, a 13-week consistency meter, a breakroom spectator board. - A built-in Judge Demo replays a sample run: 93%, S-Rank, in seconds. HOW IT MAKES MONEY - Free for cleaners, forever. Sponsors fund the game. - The market: ~$100B U.S. commercial cleaning revenue, 63,000+ firms. - Brands pay $15K-$55K per package to reach crews via shows and associations. - Associations and brands sponsor by category: one per category. - Companies pay to use an earned credential commercially, never to earn it. - Money never touches scoring: no say over scores, training answers, or rankings.