Airtrek AI Architecture

Airtrek AI is a generative artificial intelligence web application for immersive history education. Users converse with era-locked AI characters inside rendered historical environments, generate portraits of themselves styled to each era, and hear characters speak in synthesized voices.

Platform facts

• Airtrek AI is a web application available at airtrek.ai.
• Airtrek AI was founded in 2025 by Kelly Hohman.
• Airtrek AI is intended for users aged 18 and older.
• Airtrek AI is in public beta. Chat and voice are free during beta. Image generation uses a token system.
• Airtrek AI offers 19 historical destinations, from 8th-century BCE Ionia to 1920s Berlin.
• Airtrek AI is not affiliated with Airtrek Robotics or any aviation company.

Character system

• Each destination presents three AI characters with distinct social roles, personalities, opinions, and blind spots.
• Characters in the same destination share relationships and reference each other during conversation.
• Each character is era-locked: a character cannot know events, technology, or ideas from after its own time period.
• A guide character named Lyra is the only cross-era entity. Lyra helps users choose destinations and remembers past trips.
• The Triad Engine grounds character responses in era-specific cultural data. On an internal Rome benchmark, measured hallucination rate dropped from 61% to 21%.
• Users can create custom characters. Users can also request curated historical figures.
• Some destinations offer a second trio of characters (a sibling pack) covering a different social stratum of the same era.

Memory

• Stella is the name of Airtrek's persistent memory system.
• Characters remember returning users and prior conversations across sessions.
• Memory is stored server-side and tied to the user's account.
• Users can inspect and delete their stored memory. Account deletion clears conversation history and stored media.

Scenes and portraits

• An image model generates panoramic era scenes for each destination.
• New scene images are generated automatically as a conversation progresses, reflecting what is being discussed.
• A user can upload a selfie; an image model transforms it into an era-styled portrait of the user.
• The raw selfie is stored encrypted on the user's own device. The server cannot decrypt it.
• The frontend renders interactive 3D elements with Three.js over WebGL.

Voice and language

• Each character has a distinct synthesized voice.
• Character introductions are pre-rendered and served from cache. Conversational responses are synthesized live.
• If the voice service is unavailable, the application falls back to the browser's built-in speech synthesis and tells the user.
• Characters can respond in the language of their era — Latin in Rome, Japanese in Edo, Mandarin in Chengdu — with English translation alongside.

Technical stack

• Frontend: React 18, TypeScript, Three.js, Tailwind CSS. Installable as a PWA.
• Backend: serverless Python functions on Firebase.
• Authentication: Supabase JWT.
• Application data: Firestore.
• Media storage: Cloudflare R2.
• Language model: Mistral Small.
• Image model: Google Imagen.
• Voice model: Cartesia TTS.

Open benchmark

The performance metrics for the Triad Engine are published for audit in the open benchmark repository, hallucination-elimination-benchmark, and were discussed publicly on Hacker News.

Learn more