AIOT
The home that knows you — without selling you out.
Team Members
Jason Park
Dave Song
Yukti Poddar
The Problem
The "Smart" Home is Just a Remote Control
Current IoT ecosystems are fragmented and reactive. We have "smart" lightbulbs, "smart" thermostats, and "smart" appliances, yet the intelligence remains siloed within individual devices. This creates a fundamental paradox: the more "smart" devices we add to a home, the higher the cognitive load on the resident to manage them.
Today’s smart homes organize commands, not intent. To accomplish a simple transition—like moving from a high-stress workday to a relaxing dinner—a user must act as a manual orchestrator: adjusting the thermostat, selecting a playlist, dimming specific light zones, and triggering kitchen appliances. Each session starts from zero, requiring the user to rebuild their environment’s context through a series of fragmented apps and rigid voice commands.
In short: We have built homes that can hear us, but they still don't understand us.
Findings from Literature Review
Our team examined academic writing, industry talks, and design essays from researchers and practitioners thinking about the next era of AI-human interaction. Across these sources, a clear picture emerged for what an AI-native home should be: not a smarter remote control, but an environment that infers, acts within bounded autonomy, and earns trust through visible limits rather than confident outputs. Three findings shaped our concept directly.
The interface should disappear into the environment.
Doug Cook's The Prompt-Box Paradox argues that interfaces are still frozen in text input, pushing the cognitive work of articulating intent onto the user. The forward path is generative UI that appears when needed and dissolves when the task ends.
Autonomy needs explicit limits.
Current interfaces often hide internal reasoning, leading to misplaced "over-trust" and missed errors. Because hallucinations are a core property of probabilistic models, process visibility is critical. Systems must replace vague progress bars with "layered status indicators" and "slow users down" during high-stakes tasks to ensure human judgment remains the final filter.
Trust is built through visible limits, not confident outputs.
Matt Jones's BASAAP framework argues AI earns trust by being transparent about what it cannot do. For a home with cameras and microphones, transparency has to be something you can see.
Solution:
An AI home system that reads context
Haven is an AI-native home system that reads context , your sleep cycle, your stress level, your proximity to the stove, and responds before you ask. Processed entirely on a local hub in your home. No cloud. No data leaving your walls. The home adjusts. You stay in control. Nothing is sent anywhere.
Two core principles guide the design:
Design the inference, not the command.
Instead of better input surfaces such as voice commands, app taps, dashboards, Haven figures things out on its own. The best interaction is no interaction at all in this context.
Own the environment, not the social layer.
Haven acts directly on lights, scent, and temperature, low-stakes, reversible, personal. Anything touching other people gets surfaced as a notification. Haven proposes; you decide.
How it works
Haven is structured across four layers, each feeding into the next, and the last one feeding back into the first.
Layer 1
Sensing
Camera nodes, microphones, motion sensors, and an Apple Watch capture raw signals: video, audio, presence, heart rate, and proximity. All streams stay on the local hub.
Layer 2
Inference
Five parallel models read the data: a vision model for emotion and presence, an audio model for tone, a motion model for occupancy, a biometric model for stress, and a usage model that learns your patterns over time.
Layer 3
Decision
A context engine synthesizes every inference output against your local memory of habits and schedule, then decides what to do. Manual overrides skip inference entirely and write directly to this layer.
Layer 4
Actuation
The home responds, lights shift, scent changes, a single button appears on your phone, your watch taps your wrist. Every action becomes new sensing data, closing the loop. Haven learns from its own behavior, not just yours.
Key components
Four components do the heaviest design work, each one translates a literature finding into a tangible interaction surface.
Proximity UI
Your phone doesn't show a smart home dashboard. When you walk toward the stove, the stove interface appears. Near the coffee maker, one button. Step away, it clears. The interface finds you based on where you are and reconfigures around your movement through the space.
Notification & Override
When Haven wants to act on something social, it surfaces a notification that shows what it sensed, what it proposes, and a single override button. It doesn't bury the reasoning in a settings panel. It shows its work and waits.
Privacy Mode (Hardware Interrupt)
Privacy is not a setting. When you toggle Privacy Mode, a mechanical shutter physically closes over every camera node, microphones go silent at the hardware level, and the hub stops all inference. There is nothing to intercept, nothing running in the background. You can see the shutter close. You don't have to trust the company, you can verify it with your own eyes.
Behavioral Learning Layer.
Haven starts with general defaults and gradually builds a model specific to you; your routines, rhythms, and preferences. Every actuation feeds back into the data it learns from, so the system gets sharper the longer you live with it.
Demo
See it in action
A walkthrough of Haven across a single day in Alex's life from a quiet morning wake-up to an after-work wind-down to a moment where Haven pauses to ask before changing a plan that involves someone else.
Storyboard
From the top: waking up without an alarm, the kitchen anticipating his morning, coming home to a room that adjusts itself, Haven asking before acting, and the evening with Jordan.
Why It Matters
The smart home promise has existed for a decade and mostly delivered a fancier remote control. You still open the app. You still say the right words to the right device. The home is smart in the way a calculator is smart - fast, accurate, completely dependent on your input.
Haven asks a different question: what if the home already knew? As AI moves into physical space, the design opportunity is no longer better commands but better inference - environments that respond to intent rather than instruction, act within bounded autonomy, and prove their limits through hardware rather than policy. This is what it looks like when AI is woven into the home instead of layered on top of it.
Reflection
Designing Haven pushed us to think about AI not as a feature but as an inhabitant of physical space - something that lives alongside you and has to earn its place over time. The hardest problems weren't technical. They were about restraint: when the system should act, when it should ask, and when it should disappear entirely.
A key takeaway was that the real design opportunity sits in the space between sensing and action. Designing the inference, the autonomy limits, and the visible boundaries became just as important as designing any individual interface. The behavioral learning model also exposed a gap we didn't fully solve - a legibility layer that lets users see what Haven thinks it knows about them and correct it. For a system built on trust, that's the next problem worth taking on.
AIOT
The home that knows you — without selling you out.
Team Members
Dave Song
Yukti Poddar
Jason Park
The Problem
The "Smart" Home is Just a Remote Control
Current IoT ecosystems are fragmented and reactive. We have "smart" lightbulbs, "smart" thermostats, and "smart" appliances, yet the intelligence remains siloed within individual devices. This creates a fundamental paradox: the more "smart" devices we add to a home, the higher the cognitive load on the resident to manage them.
Today’s smart homes organize commands, not intent. To accomplish a simple transition—like moving from a high-stress workday to a relaxing dinner—a user must act as a manual orchestrator: adjusting the thermostat, selecting a playlist, dimming specific light zones, and triggering kitchen appliances. Each session starts from zero, requiring the user to rebuild their environment’s context through a series of fragmented apps and rigid voice commands.
In short: We have built homes that can hear us, but they still don't understand us.
Findings from Literature Review
Our team examined academic writing, industry talks, and design essays from researchers and practitioners thinking about the next era of AI-human interaction. Across these sources, a clear picture emerged for what an AI-native home should be: not a smarter remote control, but an environment that infers, acts within bounded autonomy, and earns trust through visible limits rather than confident outputs. Three findings shaped our concept directly.
The interface should disappear into the environment.
Doug Cook's The Prompt-Box Paradox argues that interfaces are still frozen in text input, pushing the cognitive work of articulating intent onto the user. The forward path is generative UI that appears when needed and dissolves when the task ends.
Autonomy needs explicit limits.
Current interfaces often hide internal reasoning, leading to misplaced "over-trust" and missed errors. Because hallucinations are a core property of probabilistic models, process visibility is critical. Systems must replace vague progress bars with "layered status indicators" and "slow users down" during high-stakes tasks to ensure human judgment remains the final filter.
Trust is built through visible limits, not confident outputs.
Matt Jones's BASAAP framework argues AI earns trust by being transparent about what it cannot do. For a home with cameras and microphones, transparency has to be something you can see.
Solution:
An AI home system that reads context
Haven is an AI-native home system that reads context , your sleep cycle, your stress level, your proximity to the stove, and responds before you ask. Processed entirely on a local hub in your home. No cloud. No data leaving your walls. The home adjusts. You stay in control. Nothing is sent anywhere.
Two core principles guide the design:
Design the inference, not the command.
Instead of better input surfaces such as voice commands, app taps, dashboards, Haven figures things out on its own. The best interaction is no interaction at all in this context.
Own the environment, not the social layer.
Haven acts directly on lights, scent, and temperature, low-stakes, reversible, personal. Anything touching other people gets surfaced as a notification. Haven proposes; you decide.
How It Works
Haven is structured across four layers, each feeding into the next, and the last one feeding back into the first.
Layer 1
Sensing
Camera nodes, microphones, motion sensors, and an Apple Watch capture raw signals: video, audio, presence, heart rate, and proximity. All streams stay on the local hub.
Layer 2
Inference
Five parallel models read the data: a vision model for emotion and presence, an audio model for tone, a motion model for occupancy, a biometric model for stress, and a usage model that learns your patterns over time.
Layer 3
Decision
A context engine synthesizes every inference output against your local memory of habits and schedule, then decides what to do. Manual overrides skip inference entirely and write directly to this layer.
Layer 4
Actuation
The home responds, lights shift, scent changes, a single button appears on your phone, your watch taps your wrist. Every action becomes new sensing data, closing the loop. Haven learns from its own behavior, not just yours.
Key components
Four components do the heaviest design work, each one translates a literature finding into a tangible interaction surface.
Proximity UI
Your phone doesn't show a smart home dashboard. When you walk toward the stove, the stove interface appears. Near the coffee maker, one button. Step away, it clears. The interface finds you based on where you are and reconfigures around your movement through the space.
Notification & Override
When Haven wants to act on something social, it surfaces a notification that shows what it sensed, what it proposes, and a single override button. It doesn't bury the reasoning in a settings panel. It shows its work and waits.
Privacy Mode (Hardware Interrupt)
Privacy is not a setting. When you toggle Privacy Mode, a mechanical shutter physically closes over every camera node, microphones go silent at the hardware level, and the hub stops all inference. There is nothing to intercept, nothing running in the background. You can see the shutter close. You don't have to trust the company, you can verify it with your own eyes.
Behavioral Learning Layer.
Haven starts with general defaults and gradually builds a model specific to you; your routines, rhythms, and preferences. Every actuation feeds back into the data it learns from, so the system gets sharper the longer you live with it.
Demo
See it in action
A walkthrough of Haven across a single day in Alex's life from a quiet morning wake-up to an after-work wind-down to a moment where Haven pauses to ask before changing a plan that involves someone else.
Storyboard
From top left: waking up without an alarm, the kitchen anticipating his morning, coming home to a room that adjusts itself, Haven asking before acting, and the evening with Jordan.
Why It Matters
The smart home promise has existed for a decade and mostly delivered a fancier remote control. You still open the app. You still say the right words to the right device. The home is smart in the way a calculator is smart - fast, accurate, completely dependent on your input.
Haven asks a different question: what if the home already knew? As AI moves into physical space, the design opportunity is no longer better commands but better inference - environments that respond to intent rather than instruction, act within bounded autonomy, and prove their limits through hardware rather than policy. This is what it looks like when AI is woven into the home instead of layered on top of it.
Reflection
Designing Haven pushed us to think about AI not as a feature but as an inhabitant of physical space - something that lives alongside you and has to earn its place over time. The hardest problems weren't technical. They were about restraint: when the system should act, when it should ask, and when it should disappear entirely.
A key takeaway was that the real design opportunity sits in the space between sensing and action. Designing the inference, the autonomy limits, and the visible boundaries became just as important as designing any individual interface. The behavioral learning model also exposed a gap we didn't fully solve - a legibility layer that lets users see what Haven thinks it knows about them and correct it. For a system built on trust, that's the next problem worth taking on.