Agent 01
Posture Agent
Watches: accelerometer · Controls: fan speed
Back sleepers need more airflow than side sleepers. Prone sleepers need the least. Real-time classification across supine, lateral, prone and fetal.
Work / AI · Agents / DreamBreeze
No. 21 · AI · Agents · Wellness
Your phone already knows
how you sleep.
Now it does something about it.
You charge your phone every night. You place it on the nightstand. It sits there for eight hours doing nothing. DreamBreeze turns those hours into the most useful part of your day. Four AI agents, zero hardware, fully on-device.
4AI agents · blackboard
0Wearables required
100%On-device inference
$411BLost productivity per year
68%Adults with thermal sleep issues
Act I · The Problem
During REM sleep, your body cannot regulate its own temperature.
You spend one third of your life asleep. That is twenty-six years. For most of those years, you are too hot, too cold, tangled in sheets, or woken by a fan running at the wrong speed.
3-5
Fan adjustments per night the average sleeper makes
Those are only the times you actually wake up. Dozens more micro-arousals you never remember.
Smart sleep tech is a $30bn category. Most of it costs $300+ in hardware. Wearables, smart mattresses, climate sleepers, sensor pads. Ninety-five percent of the world that already owns a phone gets nothing.
40+
Times per night your body's thermal needs change
A binary fan, on or off, cannot follow that curve. Nobody is solving this. Until DreamBreeze.
Act II · One Night, Watched
A full night, compressed into sixty seconds.
Place the phone on the mattress. Go to sleep. The accelerometer reads posture. Edge AI estimates the stage. Four agents negotiate. The fan, the sound, the wake-up window all adapt to the body underneath.
10:30 PM
Bedtime. Fan starts low. Pink noise fades in.
11:15 PM
You roll to your side. Fan adjusts down. Sound deepens.
12:00 AM
Deep sleep detected. Fan drops to a whisper. Brown noise.
02:30 AM
REM cycle. Thermal agent raises fan, your body cannot self-regulate.
04:00 AM
You shift to prone. Posture agent adapts instantly.
06:15 AM
Light sleep. Energy agent begins gentle wake sequence.
06:45 AM
Good morning. Cognitive readiness score: 82 / 100.
Act III · The Product
Four agents.
One blackboard.
One quiet decision per second.
DreamBreeze does not use a single algorithm. Four specialised agents collaborate over a shared blackboard. They do not take turns. They negotiate.
Agent 02
Thermal Agent
Watches: weather + circadian rhythm · Controls: temperature profile
Cross-references outdoor temperature, humidity, and your body's natural thermal dip at 3 AM. Five temperature cycling profiles, weather-aware via Open-Meteo.
Agent 03
Sound Agent
Watches: sleep stage · Controls: noise type and volume
Pink noise to fall asleep. Brown noise for deep sleep. Gradual fade during REM. Real-time synthesis via Web Audio API, no sample files, zero latency.
Agent 04
Energy Agent
Watches: Two-Process Model · Controls: wake sequence
Borbely's homeostatic Process S plus circadian Process C. Fourteen-day rolling sleep debt, cognitive readiness score, twenty-minute gentle arousal window.
Act IV · The Promise
Every byte of inference happens on your phone.
All accelerometer data, all posture classifications, all sleep stage estimations, all agent computations, all audio generation. None of it leaves the device.
Only encrypted sleep summaries are stored, and only if you opt in. The privacy vault lets you download, delete or inspect every byte we have.
- · GDPR (EU) compliant
- · DPDP Act 2023 (India) compliant
- · CCPA (California) compliant
- · Right to erasure, data portability
- · Never sold, never shared, never used for ads
Act V · Proof
It already works.
Live · dreambreeze-app.vercel.app
Production PWA, installable from any modern browser. Demo mode runs every feature simulated, with no hardware needed.
60-second night simulation
Visit /demo to watch a full night unfold in one minute. Skeuomorphic fan visualisation, LED stage indicators, soundscape morphing.
10-slide pitch deck
Guy Kawasaki structure at /pitch. Problem, market, product, agents, science, privacy, traction, ask. Investor-ready.
Grounded in peer-reviewed sleep science
Okamoto-Mizuno & Mizuno (2012) on posture thermoregulation. Borbely (1982) Two-Process Model. Ngo et al. (2013) on pink noise and slow-wave activity.
The Stack
End-to-end. Edge-first. WCAG 2.2 from line one.
- Next.js 16
- TypeScript 5.9
- React 19
- TensorFlow.js 4
- DeviceMotion API
- Web Audio API
- Tailwind v4
- Zustand 5
- Motion 12
- Recharts 3
- MQTT.js 5
- Supabase
- Open-Meteo
- PWA + Workbox
- Vercel Edge
If a phone can run a sleep agent, it can run anything.
I build edge-first AI that respects the device, the bandwidth and the person on the other side. If your product needs intelligence without hardware, talk to me.