Dementia changes lives—not only for those diagnosed, but for the families and caregivers who support them every day.
One of the most painful challenges is when a loved one begins to forget familiar faces, daily routines, or meaningful moments that once brought them joy.
As developers, we often talk about AI in abstract terms. But what if AI could truly help people maintain independence, dignity, and connection?
That question became the seed for RecallMe, a fully on-device AI companion designed to support dementia care—built with Flutter, optimized for Arm devices, and developed faster than ever thanks to Kiro, my AI-assisted coding environment.
🧠 Why Dementia? Why On-Device AI?
Over 55 million people live with dementia worldwide. But beyond the statistics are the daily struggles:
Forgetting family members
Losing track of routines
Feeling confused or afraid
Caregivers experiencing burnout
Privacy concerns around uploading sensitive photos to the cloud
Many AI apps depend entirely on cloud processing. For dementia care, that’s a problem:
Internet access isn’t guaranteed
Cloud photo uploads raise trust issues
Latency breaks user flow
Sensitive data shouldn't leave the device
So I asked a harder question:
Can we build a dementia-support AI assistant that runs entirely on an Arm-based phone—fast, private, and accessible for anyone?
RecallMe is my answer.
📱 Introducing RecallMe
A gentle, on-device AI companion designed for dementia support.
RecallMe combines multiple AI capabilities:
👥 Face Recognition (Fully Offline)
Point the camera at someone and hear:
“This is Sarah. She’s your daughter. I’m 87% confident.”
Built with ML Kit + a custom 256-dimensional embedding model.
🖼 Memory Recall
Tap any saved memory photo and ask:
“Tell me about this picture.”
The app generates a short, warm explanation tailored for dementia-friendly comprehension—then reads it aloud.
📅 Routine Management
Structured daily tasks
Smart notifications
Weekly progress charts
Caregiver PIN protection
🎙 Voice Interaction
Hands-free accessibility:
Speak naturally → get simple spoken responses.
🔒 100% Private
All photos, embeddings, routines, and conversations stay on the device.
No cloud uploads.
No tracking.
No internet required.
🧩 Under the Hood: How RecallMe Works
Built with Flutter
The UI is built entirely in Flutter with:
Provider for state management
Hive for fast local storage
A warm, dementia-friendly design system
Face Detection & Recognition
Face detection: ML Kit (Arm-optimized TFLite)
Embedding generation: Custom Kotlin algorithm using:
Color histograms
Spatial intensity grids
LBP textures
Edge gradients
Threshold ≈ 0.45 determines matches.
Memory Conversations
Azure OpenAI generates short, friendly responses based on:
photo metadata
memory tags
prior chat context
…then the app reads them aloud via native TTS.
Routine Engine
Timezone-aware notifications
Schedule logic stored as minutes-from-midnight
Weekly completion visualizations
It feels like a real care assistant—not a typical reminder app.
⚡ How Kiro Accelerated Development
Kiro became my AI engineering partner throughout the build.
🧭 Steering Documents
I defined three core documents:
product.md → dementia-friendly design guidelines
tech.md → Flutter + Kotlin + ML toolchain
structure.md → architecture rules and folder patterns
Kiro used these to generate code consistent with my vision.
✨ Vibe Coding
Instead of writing boilerplate, I asked:
“Create a routine manager screen with add/edit/delete, notifications, and weekly tracking.”
Kiro generated:
the full UI
state logic
Hive adapters
notification scheduling
Flutter navigation
What usually takes days took minutes.
Kiro didn’t just write code—it wrote code that fit perfectly into my architecture.
🛡 A Privacy-First Architecture
Everything happens offline:
Face embeddings → local
Routine logs → local
Memory conversations → local or optional Azure
Sensitive keys → encrypted storage
In dementia care, privacy isn’t optional—it’s essential.
🚀 Arm Optimization: Why On-Device AI Works
The app runs smoothly even on mid-range phones because of:
NEON SIMD vectorized loops
ML Kit’s TFLite acceleration
Big.LITTLE architecture awareness
Efficient image-processing patterns
On-device AI is not only possible—it’s powerful.
🧠 What I Learned While Building RecallMe
On-device ML can outperform cloud ML when designed efficiently
Dementia-friendly UX requires simplicity, warmth, and clarity
AI must be privacy-first—especially in healthcare
Latency is critical for elderly usability
Kiro supercharges development when guided with proper context
🔮 What’s Next for RecallMe
A fully on-device LLM (1–3B parameters, quantized)
MobileFaceNet-grade embeddings for better recognition
Multi-language voice support
Caregiver dashboard for analytics
Smart adaptive assistance
Integration with health sensors
❤️ Final Thoughts
RecallMe represents what modern AI should be:
Private
Accessible
Optimized for real devices
Built to help people—not replace them
And thanks to Flutter, Arm optimizations, and Kiro, building it became a fast, intuitive, and deeply meaningful experience.
This project is my reminder that AI isn’t just about models or performance—
it’s about improving lives.