1 of 16

老年智能手机助理

Assistify

Your AI audio companion that guides every smartphone interaction

Anthea Guo, Bonnie Wang, Dylan Chen

2 of 16

Problem

Cognitive overload

Too many icons, tiny labels, and confusing interface elements create overwhelming digital experiences.

Fear of consequences

Elderly users hesitate at every tap — they don't want to "break" anything or make costly mistakes.

Scam vulnerability

Phishing attempts and fake pop-ups specifically target the most vulnerable users.

Static tools fall short

Traditional accessibility features don't solve real-time, contextual confusion when it matters most.

3 of 16

A growing, underserved market

1.4B

People aged 60+

Worldwide population with >85% smartphone ownership in developed markets

80%

Struggle daily

Estimated elderly users who have difficulty with phone interfaces

4 of 16

Our Solution: Assistify

Multimodal Screen Understanding

ERNIE-4.5-VL analyzes your screen in real-time to understand text, icons, layouts, and contextual relationships.

Natural Voice Guidance

Speaks precise verbal instructions with no visual clutter: "At the bottom right, there's a screen icon. That activates screen sharing. Tap it once."

RAG-Powered Personalization

Vector database stores all interactions, retrieving top-5 similar past conversations to provide contextually-aware guidance that improves over time.

Example: First Time vs. Returning User

First Time: "I want to send this photo to my daughter"

Assistify: "I can see you're viewing a photo. Let's send it. What is your daughter’s name. First, tap the share icon at the bottom left—that's the small square with an arrow pointing up."

Returning User: "Send this to my daughter"

Assistify: "Sure! Let’s go through this again. Tap the share icon, then Messages, and select your daughter's contact."

5 of 16

01

Request help naturally

Elderly users simply speak their goal or question out loud

02

Continuous understanding

Assistify sees and comprehends exactly what's on screen in real-time

03

Step-by-step guidance

Clear spoken instructions for every tap, swipe, and interaction

04

Patient companion

Feels like a knowledgeable friend guiding users through any task

Assistify:

overlay-free, audio-first, multimodal assistance

6 of 16

Real-time screen understanding

ERNIE-4.5-VL anchors visual context to conversational guidance, connecting what users see with what they need to do next.

Overlay-free audio design

Verbal spatial guidance keeps the interface visually uncluttered using Apple ReplayKit no confusing overlays or additional elements to navigate.

Continuous verification

Capture → analyze → instruct → re-capture ensures accurate, adaptive guidance throughout every interaction.

RAG personalization

Learns the user's patterns and preferences over time for safer, faster, and more personalized assistance based on previous tasks.

What does Assistify do?

7 of 16

Example Screenshots

8 of 16

9 of 16

System Architecture

User Voice Input

Natural language request triggers screen capture

Background Screen Capture

1 FPS continuous monitoring, 10 evenly-sampled screenshots

ERNIE-4.5-VL Processing

OCR, icon recognition, layout analysis, state detection

RAG Context Retrieval

Query vector database, retrieve top-5 relevant past conversations

Personalized Instructions

Natural language generation with adaptive vocabulary and pacing

Screen Re-Capture & Verify

Analyze new state, compare against expected outcome, adapt next step

10 of 16

Beta Tester Independence

5 elderly testers (ages 68-82) completed tasks independently

Response Latency

End-to-end: speech input → screen analysis → speech output

Validated Use Cases

📸 Send photos to family

💳 Pay bills safely

🏥 Navigate health apps

⚠️ Detect scams in real-time

Technical Robustness

  • Cross-device compatibility
  • Battery optimization (~80MB footprint)
  • Security & privacy (TLS 1.3, isolated data)
  • Edge case coverage (crashes, multilingual)

100%

5s

Testing and Performance

92%

Task Completion

Comprehensive Testing

  • 150+ text understanding test cases
  • 200+ screenshot samples
  • 50 complex multi-step task scenarios
  • 5 elderly beta testers (ages 68-82)
  • iPhone 12-15, iOS 16-17

11 of 16

Technical Achievements & Validation

Proven Performance Across 250+ Real-World Scenarios

94.2%

Text Understanding

150 test cases: messages, healthcare, banking, social media, scam detection

89.7%

UI Classification

200 screenshot samples: buttons, fields, toggles, navigation, modals

91.2%

Personalization

50 returning user sessions validating RAG database effectiveness

88.5%

RAG Retrieval

100 query-context pairs ensuring semantic similarity

12 of 16

Engineering + product at Berkeley

Dylan Chen

Applied Math + Data Science

Systems architecture and on-device integration expertise

Anthea Guo

EECS

Product and data engineering

Bonnie Wang

EECS

Core engine development and user testing

Our team has shipped open-source projects including Promptli and Accent Classification.

We've solved hard engineering challenges like rebuilding background capture and creating seamless voice UX — and we're ready to scale.

13 of 16

Live Demo!

14 of 16

Explore Assistify

Baidu AI Studio Project

OpenAtom Demo

GitHub Repository

Video Demonstration

15 of 16

Thank You!

16 of 16