The Challenge

Perplexity AI wanted to bring their AI-powered answer engine to the living room. Samsung Smart TVs, running on the Tizen platform, reach over 50 million households globally — a massive opportunity to put conversational AI where families gather.

But TVs aren’t phones. There’s no keyboard. No touchscreen. No mouse. Users navigate with a remote control’s directional pad, and typing a question using an on-screen keyboard takes forever. The experience had to be fundamentally different.

The core constraints:

1. Voice-first interaction — Users should be able to press one button and speak their question naturally
2. Remote navigation — Every element must be accessible via up/down/left/right arrows with clear visual focus states
3. 10-foot UI — Text and controls must be readable from across a living room
4. Limited hardware — Smart TVs have significantly less processing power than phones or laptops
5. Samsung QA certification — The app had to pass Samsung’s rigorous certification process covering 100+ test cases

What Perplexity needed: A team that understood both AI integration and the unique constraints of embedded TV development.

Our Approach

We began with a week-long discovery phase, studying Tizen’s capabilities and limitations. Samsung TVs support web technologies through a customized Chromium engine, but with important restrictions: no Node.js runtime, limited memory, and strict security policies around microphone access.

The key insight: Voice input on a TV comes from external sources — the Samsung SmartThings app on a user’s phone, or a Bluetooth microphone. The Tizen platform provides device APIs to access these inputs, but they require careful permission handling and fallback strategies.

We designed the architecture around three principles:

1. Progressive enhancement — Start with keyboard input as the baseline, layer voice on top for users with compatible microphones
2. Streaming everything — AI responses stream word-by-word, giving immediate feedback even on slower connections
3. Focus management as a first-class concern — Build a complete spatial navigation system before any feature work

Technical Implementation

Custom Focus Navigation System

TV remote navigation is fundamentally different from web accessibility. Users expect:

• Press “right” → focus moves to the next logical element on the right
• Press “down” → focus moves to an element below, even if it’s in a different container
• Press “back” → return to the previous context or exit the modal

We built a custom focus management system from scratch:

Focus Architecture:
├── Page Registration (usePageFocus hook)
│   ├── Default focus element per page
│   ├── Back button behavior handlers
│   └── Focus restoration on navigation
├── Element Registration (FocusableElement components)
│   ├── Directional navigation declarations
│   ├── Visual focus ring styling
│   └── Keyboard event handling
└── Modal Layer Management
    ├── Focus trapping within modals
    ├── Stacked modal support
    └── Automatic focus restoration on close

Every focusable element declares its spatial neighbors: up="settings-button" down="mic-button" left="back-button". This explicit mapping ensures predictable navigation even when the DOM structure changes dynamically.

Real-Time Voice Recognition

For voice input, we integrated OpenAI’s Realtime API for speech-to-text. The challenge: handling the unique microphone sources available on Samsung TVs.

Microphone Priority:
1. Samsung Remote Microphone (via SmartThings app)
2. Samsung Bluetooth Microphone
3. Built-in TV microphone (limited availability)
4. Fallback: On-screen keyboard

The system automatically detects available devices, requests appropriate permissions, and provides clear user feedback throughout. When the user speaks, we show a real-time waveform visualization and display interim transcriptions as they’re processed.

Critical implementation details:

• Ephemeral key authentication — Voice sessions use short-lived tokens that refresh automatically
• Graceful degradation — If the microphone fails mid-session, we seamlessly switch to keyboard mode
• Timeout handling — Sessions automatically close after one minute of inactivity to preserve resources

Streaming AI Responses

Perplexity’s API streams responses as Server-Sent Events. On a TV, where users are watching from 10 feet away, this streaming behavior is even more important — they can start reading immediately rather than waiting for a complete response.

We implemented:

• Chunk timeout detection — If no data arrives for 10 seconds, we surface an error rather than leaving users waiting indefinitely
• Abort controller integration — Users can press “back” mid-response to cancel and ask a different question
• Citation deduplication — Source URLs are deduplicated by root domain for cleaner display
• Markdown rendering — Responses render as formatted text with proper heading hierarchy

Internationalization

Samsung required support for multiple languages from launch. We built a complete i18n system using svelte-i18n:

• Language detection — The app detects the TV’s system language and maps it to supported locales
• CMS-driven content — Promotional materials and onboarding copy are fetched from a remote CMS, allowing updates without app redeployment
• Regional dialect support — Korean (ko-KR) and German (de-DE) handle regional variations

The language system also controls the voice recognition model — sending the correct language code to OpenAI ensures accurate transcription across languages.

What Made This Work

1. TV-first architecture. We didn’t adapt a mobile or web app to TV. We built for the 10-foot experience from day one: large touch targets, spatial navigation, voice as the primary input.

2. Rigorous focus testing. We tested every screen flow with only arrow keys and the Enter button before any mouse interaction. Focus bugs that would be minor annoyances on web become app-breaking on TV.

3. Defensive error handling. TVs can’t “just refresh the page.” Every error state has a recovery path, every timeout has user-facing feedback, every microphone failure gracefully falls back to keyboard.

4. Samsung QA preparation. We studied Samsung’s certification requirements early and built compliance into our architecture — proper privilege declarations, accessibility labels, memory management patterns.

Results

The app passed Samsung’s QA certification and launched globally:

• 50M+ potential users across Samsung’s Smart TV install base
• Sub-2-second latency from button press to voice recognition start
• Multi-language launch with English, German, and Korean support
• 98%+ crash-free sessions across diverse hardware generations
• Continuous deployment pipeline with separate release, develop, and test channels

The application demonstrates that conversational AI can work beautifully on non-traditional platforms when designed for the medium’s unique constraints.

Technical Summary

Component	Technology	Purpose
Frontend	SvelteKit 2, TypeScript, Tailwind CSS	TV-optimized UI with custom focus system
Voice Recognition	OpenAI Realtime API	Real-time speech-to-text with streaming
AI Integration	Perplexity API, Server-Sent Events	Streaming AI responses with citations
Platform	Tizen SDK, Tizen Web APIs	Samsung TV integration, microphone access
Build System	Vite, Static Adapter	Optimized bundle for TV deployment
Testing	Vitest, JSDOM	Component and integration testing
Deployment	Tizen CLI, Custom Release Pipeline	Multi-channel releases (release/develop/test)