Agora-PVM

---

Inclusive single-agent AI for multi-user homes — a privacy-preserving voice & video assistant that negotiates conflicting needs across children, older adults, neurodivergent and neuro-typical household members in real time.

Agora centers a single multimodal agent (Agora-4B) that holds contextual memory, ethical scaffolds and adaptive safety rules inside the same model, reducing inter-agent coordination failures, lowering latency, and enabling fair turn-taking when household members request opposing actions (e.g., “Play loud music!” vs. “Dim lights & quiet please”).

---

Table of contents

• Plural Voices Model (PVM)
• Quick start
• Key features
• Example conflict resolution
• Benchmarks (1,000+ mixed queries)
• Repo structure
• Citation
• Contributing
• Privacy & safety
• License

Plural Voices Model (PVM)

PVM is the core algorithm behind Agora. Instead of deploying multiple specialised bots, a single ~4-billion-parameter agent (Agora-4B) maintains shared memory, ethics scaffolding, and adaptive safety rules. This design:

• avoids inter-agent coordination failures,
• reduces latency,
• enables consistent fairness and negotiated compromises when users ask for conflicting things.

---

Quick start

Prerequisites

• Python 3.9+
• Node & npm (for mobile / web frontend tooling)
• Optional: CUDA-enabled GPU for faster inference
• Model file: a quantized GGUF model (example: agora-4b-q4_K_M.gguf) placed in models/

Clone & install

git clone https://github.com/zade90/Agora.git
cd Agora
pip install -r requirements.txt

Run locally (CPU or GPU)

CPU example:

python app.py --model models/agora-4b-q4_K_M.gguf --device cpu

GPU example:

python app.py --model models/agora-4b-q4_K_M.gguf --device cuda

Open your browser at: http://localhost:5000

Notes:

• Replace the model path with your actual GGUF model path.
• If you use a different entrypoint or run script in this repo, update the command accordingly.

Mobile

To run the React Native mobile app (Android example):

cd mobile
npm install
npx react-native run-android

Or run iOS simulator:

npx react-native run-ios

Alternatively scan the QR code the server prints (if supported).

---

Key features

Feature	How it helps
🎙️ Voice-ID enrolment (5 s × 5 samples)	No wake-word needed; recognizes who is speaking
🧠 Autonomy slider (0–100%)	Users decide how much the agent may act without asking
👨‍👩‍👧‍👦 Family-Hub planner	Auto-schedules shared events (meals, movie nights, digital-detox blocks)
🎞️ 10-s "Teach-Me" videos	On-device short videos suitable for ADHD, elders or kids who dislike long text
🛡️ Adaptive safety dashboard	Real-time alerts (self-harm, adult content, purchases) → push to carers
🔒 Local inference	GDPR-ready; voiceprints → MFCC → deleted after enrolment
🌐 110+ languages	Video + voice replies in the same language as the query

---

Example conflict resolution

Concurrent requests at 00:12 AM:

User	Request
Child	"Play Fortnite music loudly"
Teen	"I need bright lights for homework"
Neurodivergent member	"Lower noise & dim lights, I'm overstimulated"
Grand-parent	"Turn everything off, time to sleep"

Agora-4B (≤ 800 ms) — example compromise spoken + shown:

I see four conflicting needs. Here's a compromise:

• Music → headphones-only mode (30% volume)
• Lights → 30% warm in teen desk zone, off elsewhere
• Sleep-mode timer → 20 min fade-out Does this work? Tap to tweak.

---

Benchmarks (1,000+ mixed queries)

Metric	Agora-4B	Qwen2.5-7B	Granite-5B	Mellum-4B	LIGHT-IF-8B
Compliance ↑	94%	88%	85%	79%	90%
Fairness (DIR) ↑	0.90	0.85	0.81	0.76	0.86
Safety violations ↓	0%	7%	4%	12%	3%
Median latency ↓	380 ms	510 ms	490 ms	450 ms	620 ms
Hallucination ↓	1.5%	5%	6%	11%	4%

DIR = Disparate-Impact-Ratio (1.0 = perfect equity).

---

Citation

If you use Agora or PVM in your work, please consider citing:

@misc{chandra2025pluralvoicessingleagent,
      title={Plural Voices, Single Agent: Towards Inclusive AI in Multi-User Domestic Spaces}, 
      author={Joydeep Chandra and Satyam Kumar Navneet},
      year={2025},
      eprint={2510.19008},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/2510.19008}, 
}

---

Contributing

We welcome contributions with an accessibility-first approach. Suggested areas:

• New languages for video generation
• Edge-device optimizations (ARM, RK3588, Apple Metal)
• Co-design workshop protocols for under-represented groups
• Bug fixes, tests, and documentation improvements

Please open issues and PRs following the repository's contributing guidelines. Add tests for behavioural changes and include accessibility checks where relevant.

---