Project Tango

Project Tango adapts the forked AURA LiveKit voice interface into a Schubert-hosted, persona-driven AI companion. The frontend keeps the orb-style WebRTC session UI, while the backend routes all LLM traffic through the existing LiteLLM proxy on Schubert.

The canonical deployment repo is the private theonlygeranium/Project-Tango repository. This public theonlygeranium/AURA repo is retained as the fork baseline and historical adaptation source.

Architecture

• frontend/ - Next.js 15 LiveKit client on port 3006.
• backend/ - FastAPI token/session API plus persona-aware LiveKit voice worker on port 8030.
• deploy/ - systemd units plus the API-only Caddy append block.
• docs/ - Codex agent constraints and bootstrap plan.

Project Tango assumes these Schubert services already exist:

• polyglot-litellm.service at http://localhost:4000
• ollama.service with qwen3.6:latest
• postgresql@18-main.service
• caddy.service
• cloudflared.service

Do not deploy new LiteLLM or Ollama services for this repo. LLM calls go through http://localhost:4000 only, authenticated with LITELLM_MASTER_KEY.

Schubert Notes

• project-tango.schubert.life already proxies to frontend port 3006; do not create a second frontend Caddy block.
• deploy/Caddyfile.tango-api is the only new Caddy append block and routes tango-api.schubert.life to backend port 8030.
• Live inspection on 2026-06-22 showed 127.0.0.1:8010 is permanently owned by Docker container asr-gateway; Tango uses 8030, which must remain free for the backend. Run deploy/schubert-preflight.sh before installing.
• Do not add WRITER_API_KEY or PALMYRA_API_KEY to Tango env files. LiteLLM already owns downstream provider credentials.
• Runtime credentials stay in /opt/Project-Tango/.env on Schubert and are not tracked in Git.

Personas

Persona	Display Name	Voice ID	LiteLLM Alias
Therapy	Damian	QF9HJC7XWnue5c9W3LkY	local/qwen3-fast
General Info	Chris (British)	HfRP3cIhYLmeNHeTvkWK	writer/palmyra-x5-voice
Meditation	Nathaniel	pFQStpMdprGFILRDrWR2	local/qwen3-fast
Pinoy Pride	Tita	smYFzUb4yrSqprnml7n5	local/qwen3-fast

The frontend sends the selected persona to /api/connection-details. The backend encodes that selection into LiveKit token metadata and room naming so the worker can load the correct system prompt, ElevenLabs voice, and LiteLLM model.

Local Setup

Backend:

cd backend
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
uvicorn main:app --host 127.0.0.1 --port 8030 --reload

Frontend:

cd frontend
npm install
cp .env.example .env.local
npm run dev -- --port 3006

Open http://localhost:3006.

Validation

1. cd backend && uvicorn main:app --host 127.0.0.1 --port 8030 --reload
2. cd frontend && npm run dev -- --port 3006
3. Confirm the persona selector shows all four personas.
4. Select Damian and connect.
5. Confirm Deepgram Nova-3 interim captions are visible.
6. Confirm ElevenLabs Flash v2.5 playback and speaking animation.
7. Confirm backend logs show http://localhost:4000, not api.openai.com.
8. Confirm Therapy uses local/qwen3-fast.
9. Switch to Chris and confirm model routing changes to writer/palmyra-x5-voice.
10. Confirm deploy artifacts exist in deploy/.
11. Run bash -n deploy/schubert-preflight.sh.

Read docs/AGENTS.md before making further changes.