One tap. A full conversation.
No typing, just talk. The agent never makes things up and shows its sources for every answer.
Mongolian speech recognition
The Omni model turns speech into text, with a vocabulary of brand names like Gerege, Gepay and eID.
Knowledge base · RAG
Finds the most relevant sections in knowledge/*.md and cites them in every answer.
Small, fast SLM
A mixture-of-experts model with 3B active parameters, picked after benchmarking six models.
Streaming TTS
Each sentence is voiced in parallel and played in order as soon as it is ready. 10 voices.
Voice cloning
Create a consented brand voice from a 5–20 second recording with Voicebox.
Transcript export
Every conversation is saved and can be exported as Markdown, TXT or JSON.
Four steps. Two seconds.
Fully modular: models, voices and the knowledge base can each be swapped independently. The API key lives only on the server.
STT
- 16 kHz WAV input
- Brand vocabulary
- Auto-send on silence
RAG
- Chunked by heading
- Disk cache
- Top-4 cosine
SLM
- 3B active parameters
- Declines instead of guessing
- 10-turn memory
TTS
- Parallel per sentence
- Streamed in order
- Tap to interrupt
No waiting. Just talking.
Measured on 2026-11-03. The first sentence is already playing before the full answer has been written.
In your pocket. By voice.
The agent ships as a native iOS app through TestFlight. If you have an invite, install it from the TestFlight app.
Gerege Voice
- Install TestFlight from the App Store
- Tap View in TestFlight in the invitation email
- Install, open the app and tap the mic to ask
No invite yet? Contact the Gerege Systems team. SwiftUI · iOS 17+
Three days. A complete product.
Until the hackathon deadline: 2026-11-06, 16:00 (UTC+8).
Foundation
- HLA and pipeline
- RAG, SLM, TTS
- iOS app on TestFlight
- Transcript export
Hardening
- Expand knowledge base
- On-device testing
- Beta App Review
Quality
- Consented brand voice clone
- Barge-in
- 30-question evaluation
Ship
- 12:00 code freeze
- Demo video
- 16:00 submit