Chronos: The Living Archive

Talk to history. An AI persona narrates the past through real-time voice (Gemini Live API) while generating contextual images inline. Interrupt, ask, explore — history has never felt this alive.

Comment

Inspiration

History education has always been static — textbooks, documentaries, and lectures. I asked: What if you could actually talk to someone who lived through history? What if a soldier from Gallipoli could describe the trenches in his own voice, while AI-generated images of the scene appeared on your screen?

That question led to Chronos: The Living Archive — an app that transforms history from something you read into something you experience.

What it does

Chronos lets users select an AI persona (a soldier, a professor, or a time traveler) and a historical era, then engage in real-time voice conversations powered by the Gemini Live API. As the AI narrates events:

  • AI-generated images appear inline, visualizing the scenes being described
  • Users can interrupt anytime to ask questions (X-Factor feature)
  • Text narration flows alongside voice and images for a rich multimedia feed
  • Full English and Turkish bilingual support

Three Personas

Persona Style
The Witness Emotional, first-person — a soldier who lived through events
The Professor Academic, analytical — a historian providing scholarly analysis
The Time Traveler Cinematic, immersive — transports you to the scene

Three Historical Eras

  • Fall of Constantinople (1453) — The Ottoman siege that changed the world
  • Battle of Gallipoli (1915) — WWI's most devastating campaign
  • D-Day: Normandy Landings (1944) — The largest seaborne invasion in history

How we built it

Frontend: Flutter (Android) with Riverpod state management. A custom Android Platform Channel (Kotlin) captures microphone audio at 16kHz PCM. Audio playback uses flutter_pcm_sound at 24kHz.

Voice AI: Gemini Live API via WebSocket (BidiGenerateContent) using gemini-2.5-flash-native-audio-latest for real-time bidirectional audio streaming with Voice Activity Detection.

Image Generation: Three-model cascading fallback system:

  1. gemini-2.5-flash-image (primary)
  2. gemini-2.0-flash-exp-image-generation (fallback)
  3. imagen-4.0-fast-generate-001 (fallback)

All image generation runs in a background Dart Isolate (Isolate.run()) to prevent audio crackling during HTTP requests. Auto-timer triggers contextual images (first at 8s, then every 35s).

Backend: FastAPI on Google Cloud Run using the Google GenAI SDK (google-genai). Provides /generate-image, /health, and /models endpoints. Deployed via Cloud Build with automated deploy.sh script.

Infrastructure: Google Cloud Run + Container Registry + Cloud Build. Docker container with python:3.12-slim base image.

Challenges we ran into

  1. Audio crackling during image generation: HTTP requests and base64 decoding on the main isolate caused audio dropouts. Solved by moving the entire image pipeline to Isolate.run().

  2. Image API rate limits: Free-tier API keys returned limit: 0 for image generation. Solved by upgrading to Tier 1 and implementing a cascading fallback across three different models.

  3. Voice Activity Detection timing: Getting natural conversation flow with Gemini Live API required careful tuning of silence detection and end-of-turn signals.

  4. Bilingual persona prompts: Crafting system prompts that make AI personas speak naturally in both English and Turkish while maintaining historical accuracy.

Accomplishments that we're proud of

  • Seamless voice + image experience: Users can have a natural conversation while AI-generated historical images appear inline.
  • Interrupt feature (X-Factor): One-tap interrupt that clears the audio buffer and auto-opens the microphone, creating truly natural conversation
  • Background Isolate architecture: Too little audio artifacts during image generation thanks to complete isolation of network calls
  • Three distinct personas: Each with unique narrative styles that fundamentally change the storytelling experience

What we learned

  • Gemini Live API's WebSocket protocol enables remarkably natural voice interactions with proper VAD tuning
  • Dart Isolates are essential for CPU-intensive or network-heavy tasks in Flutter when audio is playing
  • Cascading model fallbacks provide resilience — if one model is unavailable, the experience degrades gracefully with SceneCard fallbacks
  • The Google GenAI SDK simplifies server-side integration significantly

What's next for Chronos

  • More historical eras: Ancient Rome, the Moon Landing, the Renaissance
  • User-created eras: Let users define custom historical scenarios
  • AR integration: Overlay historical scenes onto real-world locations
  • Classroom mode: Multi-user sessions for history education
  • iOS support: Expand beyond Android

Built With

  • cloudrun
  • dart
  • docker
  • fastapi
  • flutter
  • geminiliveapi
  • genai
  • imagen
  • kotlin
  • riverpod
Share this project:

Updates

Submission history