Designing Human-AI Interaction for
Preventative Pet Health
I led the end-to-end design of Felis, an AI-powered biosensing system that helps pet owners identify subtle health changes before they become visible symptoms.
The core challenge was translating continuous, ambiguous behavioral data into guidance owners could act on, without allowing the AI to diagnose, anthropomorphize, or overstate what the sensors could know.
I created a product ecosystem that learns each cat's personal baseline and turns multimodal data into explainable signals, enabling owners to recognize meaningful change and know when to involve a veterinarian.
There are 76.3 million pet cats in the US. Nearly all of them share the same problem: cats are biologically wired to conceal pain and vulnerability.
By the time a cat displays an obvious symptom, an underlying condition may already be advanced.
A multimodal biosensing system, a smart collar and companion app, that translates everyday cat routines into opportunities for preventative care.
It observes patterns, learns what is normal for each cat and identifies sustained departures from the baseline.
The experience is designed around four questions.
The home screen provides a concise view of the cat's current activity and whether it is consistent with its usual behavior.
A live point-of-view feed offers environmental context, while behaviorally grounded state labels provide immediate orientation without making emotional claims.
The Day at a Glance timeline turns continuous behavioral data into a readable daily sequence.
Owners can scan periods of activity, rest, eating, drinking, litter-box use, vocalization, and relevant point-of-view clips without reviewing raw metrics or hours of footage.
Signal-detail views compare recent behavior with the cat's established baseline.
Instead of showing generic population ranges, Felis explains how the current pattern differs from what is normal for that specific cat, including its duration, magnitude, and related signals.
The Companion AI helps owners understand the evidence behind a detected change.
It can explain what happened, connect it to the cat's personal history, and clarify what the sensors cannot determine. When a sustained pattern deserves attention, it recommends a veterinary conversation without proposing a diagnosis.
All of this is made possible by a smart collar that continuously captures behavioral and physiological signals throughout a cat's daily life.
By combining multiple sensing modalities, the collar provides the foundation for Felis's ability to detect meaningful changes, establish personal baselines, and surface explainable insights over time.
Designed for adult cats.
Felis does not translate raw sensor data directly into health conclusions. Its information architecture moves through several layers.
THE DECISIONS BEHIND FELIS
The final experience emerged from a series of research findings, failed assumptions, and deliberate product constraints.
Starting with the wrong interaction model
My first concept, Iris, used AI-powered posture analysis to interpret a cat's body language through a phone camera.
Testing revealed that it was solving the wrong moment.
- Owners opened the tool only after something appeared wrong, making the interaction reactive.
- It also depended on the owner being present, even though many meaningful behavioral shifts happened overnight or when no one was watching.
This pivot transformed Felis from a camera tool into an ambient product ecosystem.
Across 11 sessions with pet owners and animal-behavior experts, three findings shaped the system.
Population norms were not personal enough.
Owners missed gradual drift.
Existing products overclaimed or undersolved.
The meaning of a behavior depended on whether it represented a change for that individual cat.
Owners noticed acute events but struggled to recognize small changes accumulating across days or weeks.
Some products overinterpreted sensor data while others showed raw activity without explaining what it meant.
I designed a delta-first architecture that measures present behavior against the cat's own history.
The system considers:
- Whether a pattern differs from normal
- How large the change is
- How long it has persisted
- Whether related signals are also changing
This made Felis personally relevant without relying on diagnostic claims.
Research into animal perception challenged the assumption that human emotional labels could accurately describe a cat's internal experience.
I converted that principle into a language rule:
Felis may describe observable behavior, patterns, and deviations. It may not assign an emotional state that its sensors cannot verify.
Instead of:
Kimchi seems anxious.
Felis says:
Kimchi has slept four hours less than her usual average this week.
The interface preserves the distinction between what the system observed and what a person may infer from it.
A language model could associate a behavioral pattern with a possible medical condition. I chose not to design that interaction.
Felis can detect that litter-box visits have increased. It cannot determine why.
Jumping from behavioral evidence to medical cause would cross the boundary between sensing and veterinary judgment. It could also make an owner feel falsely reassured that the information loop had been closed.
I defined the Companion AI around five steps:
- Identify the change.
- Explain the supporting evidence.
- Communicate uncertainty.
- Recommend an appropriate next step.
- Stop before diagnosis.
This allows Felis to support action without impersonating clinical authority.
Felis shifts pet monitoring from episodic observation to ambient, longitudinal awareness.
Instead of asking owners to identify every meaningful moment themselves, the system notices patterns over time and surfaces only the changes that may deserve attention.
Its value is not in claiming to know whether a cat is ill.
Its value is in making subtle change visible early enough for someone to act.
Kimchi, my ginger tabby, is the reason Felis exists.
Atisha + Kimchi
The project began with the fear of recognizing a health problem too late. It became an exploration of a broader product question:
How should AI communicate when the available evidence is meaningful but incomplete?
Designing Felis taught me that trustworthy AI is not created by making a system sound more certain.
It is created by making the boundaries of its knowledge visible, and still helping people decide what to do next.