---
title: JARVIS - Personal AI Agent
emoji: 🤖
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: "4.26.0"
app_file: app.py
pinned: true
license: mit
tags:
  - conversational
  - personal-assistant
  - llm
  - phi-3
  - jarvis
  - voice
  - memory
  - agentic
hardware: cpu-upgrade
---

# J.A.R.V.I.S — Just A Rather Very Intelligent System

> *"Sometimes you gotta run before you can walk."*

Your personal AI assistant — conversational, voice-enabled, and memory-powered. Runs locally on a HuggingFace Space with **16 GB RAM** and no GPU required (thanks to 4-bit quantisation).

---

## ✨ Features

| Feature | Detail |
|---|---|
| 🧠 **Local LLM** | Microsoft Phi-3-mini-4k-instruct (3.8B params, 4-bit ≈ 2-4 GB RAM) |
| 💬 **Chat interface** | Clean Gradio UI — type or use voice |
| 🎙️ **Voice replies** | Edge-TTS (Microsoft GuyNeural, no GPU, minimal cloud) |
| 🗂️ **Persistent memory** | SQLite facts + conversation history |
| 💡 **Thinking engine** | Background anticipation — JARVIS predicts your needs |
| 🔧 **Commands** | `status` · `remember X is Y` · `anticipate` |

---

## 🚀 Quick Start (local)

```bash
git clone https://huggingface.co/spaces/<your-username>/jarvis
cd jarvis
pip install -r requirements.txt
python app.py
```

Open `http://localhost:7860` in your browser.

---

## ⚙️ Environment Variables

| Variable | Default | Description |
|---|---|---|
| `JARVIS_MODEL` | `microsoft/Phi-3-mini-4k-instruct` | HuggingFace model ID |
| `JARVIS_QUANT` | `4bit` | Quantisation: `4bit` / `8bit` / `none` |
| `JARVIS_MEMORY_DIR` | `./data/memory` | Where SQLite DB is stored |
| `JARVIS_TTS_VOICE` | `en-US-GuyNeural` | Edge-TTS voice name |
| `JARVIS_THINKING` | `true` | Enable background thinking engine |

Set in HuggingFace Space **Settings → Variables & Secrets** or in a `.env` file locally.

---

## 💾 RAM Requirements

| Model | Quant | RAM |
|---|---|---|
| Phi-3-mini-4k-instruct (3.8B) | 4-bit | ~3–4 GB |
| Phi-3-mini-4k-instruct (3.8B) | fp16 | ~7.6 GB |
| Mistral-7B-Instruct-v0.2 | 4-bit | ~4–5 GB |
| Mistral-7B-Instruct-v0.2 | fp16 | ~14 GB |

**Recommended for 16 GB RAM:** Phi-3-mini 4-bit (fast) or Mistral-7B 4-bit (deeper reasoning).

---

## 🗂️ Project Structure

```
app.py          — Gradio UI + entry point
brain.py        — HuggingFace model inference (lazy-loaded)
memory.py       — SQLite facts + conversation history
thinking.py     — Background anticipation engine
voice_tts.py    — Edge-TTS text-to-speech
config.py       — All settings via env vars
requirements.txt
README.md       — This file (HuggingFace model card)
```

---

## 🤝 Credits

Built on top of:
- [Microsoft Phi-3](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct)
- [HuggingFace Transformers](https://github.com/huggingface/transformers)
- [Gradio](https://gradio.app)
- [Edge-TTS](https://github.com/rany2/edge-tts)
- Original JARVIS project — OpenJarvis architecture inspiration