Jazari

A Turkish LLM, a voice coach, and a mobile app — named after the 12th-century engineer from Diyarbakır who built the world's first programmable automata.

Why "Jazari"

Al-Jazari (1136–1206) built programmable automata, water clocks, and mechanical servants 800 years ago in Diyarbakır. Systematic, precise, ingenious. The name felt right for an AI that tries to be the same.

The Ecosystem

Jazari 4B SFT TR — Fine-tuned Turkish language model, published on HuggingFace. Trained on RTX 5070 Ti with MLX. 4 billion parameters, supervised fine-tuning on curated Turkish data.

Jazari AI Life Coach — Voice-first coaching agent using Gemini Live API with real-time bidirectional audio. 5 specialist sub-agents for career, health, finance, discipline, and memory. Uses the Ebbinghaus forgetting curve — identity facts persist for months, casual observations fade in days. Submitted to the Gemini Live Agent Challenge.

Jazari Mobile — Android app on Google Play under Jazari Labs developer account. Currently in closed beta testing.

Challenges

Voice mode can't call tools. Gemini Live API's run_live() doesn't support function calling. Workaround: voice transcripts auto-save to memory when the session ends, so nothing discussed in voice mode is lost.

Ephemeral containers. LanceDB data disappears when Cloud Run containers restart. Solution: Firestore backup on every write, full restore on startup. Memory persists across deployments.

Embedding model migration. The Gemini embedding API changed during development. Migrating from the old model to gemini-embedding-2-preview required re-embedding all existing memories — a one-time cost but a reminder to abstract the embedding layer.

Stack

Gemini 2.5 Flash · Google ADK · LanceDB · FastAPI · Cloud Run · Firestore · React · HuggingFace · MLX · PyTorch