DEV Community

Esther Studer
Esther Studer

Posted on

I Built an AI Pet Therapy Companion in Python — Here's What I Learned About Human Emotion

#ai #python #showdev #webdev

There's a dirty secret in mental wellness tech: most apps ignore the most powerful therapeutic tool humans have had for 15,000 years — their pets.

I spent the last 6 months building an AI layer on top of pet-assisted therapy data. What I found changed how I think about emotional AI entirely.

The Problem With Wellness Apps

Every burnout-prevention, anxiety-tracking, mood-journaling app follows the same playbook:

  1. Ask the user how they feel
  2. Show them a graph
  3. Suggest meditation

It doesn't work long-term. Why? Because it's self-referential. You're asking a depressed brain to accurately report on its own depression.

Pets short-circuit this entirely.

What Dogs Know That Algorithms Don't

A dog doesn't ask you how you're doing. It detects it — through micro-movements, scent, heart rate variability, breathing rhythm. Then it acts.

I started wondering: can we build that detection layer artificially?

import anthropic
import cv2
import mediapipe as mp

def analyze_user_state(video_frame):
    """
    Rough sketch: use pose + facial landmarks to infer emotional state
    then route to appropriate pet-companion response
    """
    mp_face_mesh = mp.solutions.face_mesh
    mp_pose = mp.solutions.pose

    with mp_face_mesh.FaceMesh(static_image_mode=True) as face_mesh:
        results = face_mesh.process(cv2.cvtColor(video_frame, cv2.COLOR_BGR2RGB))

        if not results.multi_face_landmarks:
            return {"state": "unknown", "confidence": 0.0}

        landmarks = results.multi_face_landmarks[0]

        # Eye aspect ratio for stress detection
        # Brow furrow distance for anxiety
        # Mouth corner deflection for mood
        features = extract_facial_features(landmarks)

    return classify_emotional_state(features)

def classify_emotional_state(features):
    client = anthropic.Anthropic()

    prompt = f"""
    Given these facial feature measurements:
    - Eye openness: {features['eye_ar']:.3f} (1.0 = wide open)
    - Brow tension: {features['brow_distance']:.3f} (lower = more furrowed)
    - Lip corner angle: {features['lip_angle']:.1f} degrees

    Classify the emotional state and suggest a pet-companion interaction:
    Return JSON: {{"state": str, "intensity": 0-10, "suggested_interaction": str}}
    """

    response = client.messages.create(
        model="claude-3-5-sonnet-20241022",
        max_tokens=256,
        messages=[{"role": "user", "content": prompt}]
    )

    return json.loads(response.content[0].text)
Enter fullscreen mode Exit fullscreen mode

This is a simplified sketch — production needs proper privacy handling, consent flows, and on-device processing. But the concept holds.

The Surprising Results

When I ran this against session data (anonymized, consented), the AI companion responses that worked best weren't the ones that addressed the emotional state directly.

They were the ones that redirected attention outward — exactly what a dog does when it drops a toy in your lap.

INTERACTION_MATRIX = {
    "anxious": [
        "Your virtual companion nudges you with a tennis ball 🎾",
        "Time for a 5-minute 'walk' — step outside, we'll wait",
        "[Pet name] is doing the head-tilt thing. Look."
    ],
    "low_energy": [
        "Lie on the floor for 2 minutes. Seriously. It works.",
        "Your companion wants belly rubs. Who's really winning here?"
    ],
    "stressed": [
        "60-second breathing exercise — your pet's watching your chest rise",
        "Name 3 things your pet would find interesting right now"
    ]
}
Enter fullscreen mode Exit fullscreen mode

Low-tech responses. High emotional resonance.

Why This Matters for AI Developers

We tend to over-engineer emotion AI. We want sentiment models, voice analysis, wearable integration.

But the insight from pet therapy research is brutally simple:

Consistent, non-judgmental presence beats sophisticated intervention.

Your AI companion doesn't need to be a therapist. It needs to be a dog.

The Architecture That Worked 

User State Detection
        ↓
Emotional Routing Layer  
        ↓
Response Selection (NOT advice-giving)
        ↓
Attention Redirection
        ↓
Micro-engagement Loop (2-5 min max)
Enter fullscreen mode Exit fullscreen mode

No therapy. No diagnosis. No "have you tried mindfulness?" Just presence, playfulness, and pattern interruption.

What's Next

I'm continuing this research with the team at MyPetTherapist — a platform that connects the science of animal-assisted therapy with AI-augmented wellness tools. If you're building in this space or just curious about the intersection of pet behavior research and emotional AI, worth a look.

The takeaway for devs: Before you reach for a complex ML pipeline, ask — what would a dog do? Sometimes the answer is "drop a tennis ball and wag." That's valid UX.

What's the simplest intervention you've shipped that had the biggest emotional impact? Drop it in the comments.

Top comments (0)