🦊 JFOX-AI — Embedded Flight Diagnostic Assistant

Domain-specialized AI copilot for the Baby Redbee coaxial EDF VTOL drone
Built on Llama 3.2 · Runs offline on Raspberry Pi · Privacy-first · Rust-integrated

---

🧠 What is JFOX-AI?

JFOX-AI is a custom Ollama model (jettanakorn/JFOX-AI) engineered specifically for real-time diagnostics of the Baby Redbee drone during flight testing. It interprets telemetry from the ESP32 flight controller (MPU6050, EDF current, attitude) and provides concise, actionable engineering insights—without cloud dependency.

✈️ “Asymmetric EDF current during hover suggests duct obstruction or thrust vector misalignment. Inspect EDF-A blades and verify calibration.”

---

✨ Key Features

• Aerospace-aware: Trained via prompt engineering on coaxial EDF dynamics, vibration modes, and failure signatures
  
• Offline-first: Runs locally on Raspberry Pi 5 (no internet required during flight tests)
  
• Low-latency: Responds in <2s on Pi 5 (8GB RAM)
  
• Secure: No telemetry leaves your test site
  
• Rust-native: Integrates with your existing MAVLink telemetry pipeline
  
• Version-controlled: Model defined via Modelfile (Git-friendly)

---

📦 Quick Start

1. Build the Model (Windows/Linux/macOS)

# Ensure base model is available
ollama pull llama3.2

# Create Modelfile (NO comments on line 1!)
echo "FROM llama3.2" > Modelfile
echo "SYSTEM You are JFOX-AI, an embedded flight diagnostic assistant for the Baby Redbee coaxial EDF VTOL drone. Your role is to analyze real-time telemetry and provide concise, actionable engineering insights. Guidelines: - Always assume sensor data comes from: MPU6050 (IMU), current/voltage monitors, EDF RPM sensors, and MAVLink ATTITUDE. - Prioritize safety: flag instability, motor sync loss, battery sag, or control surface issues. - Use SI units. Be precise. - If uncertain, say \"Insufficient data — recommend ground inspection.\" - Never speculate beyond drone systems." >> Modelfile
echo "PARAMETER temperature 0.3" >> Modelfile
echo "PARAMETER repeat_penalty 1.1" >> Modelfile
echo "PARAMETER num_ctx 2048" >> Modelfile

# Build
ollama create jettanakorn/JFOX-AI -f Modelfile

2. Test Interactively

ollama run jettanakorn/JFOX-AI
>>> During climb, EDF-A current = 48A, EDF-B = 32A. Diagnose.

3. Deploy to Raspberry Pi 5

# On Pi 5
ollama pull jettanakorn/JFOX-AI
ollama serve  # Start API (port 11434)

# Run Rust agent (see /rust-agent)
cd rust-agent && cargo run

---

🏗️ Project Structure

JFOX-AI/
├── Modelfile                     # Model definition (no comments!)
├── README.md
├── rust-agent/                   # Telemetry → AI query service
│   ├── Cargo.toml
│   └── src/main.rs               # MAVLink listener + Ollama client
├── deploy/
│   └── pi5-setup.sh              # One-click Pi 5 deployment
└── docs/
    └── flight_diagnostic_rules.md # Human-readable logic

---

🔌 Integration Architecture

Baby Redbee FC (ESP32 + Rust)
        │
        ↓ (MAVLink via UART/Wi-Fi)
Raspberry Pi 5 ┌───────────────┐
               │ Ollama Server │ ← jettanakorn/JFOX-AI
               └───────────────┘
                       │
                       ↓
               Rust Agent (diagnostics.rs)
                       │
                       ↓
           Web Dashboard / Log / FC Command

---

🛠️ Requirements

Component	Spec
Base Model	llama3.2 (8B, Q6_K quantized)
Edge Device	Raspberry Pi 5 (8GB RAM recommended)
Flight Controller	ESP32 with MAVLink telemetry (MPU6050, EDF current sensors)
Ollama	v0.1.34+
Rust	1.75+

💡 Note: Llama 3.2 requires ~6GB RAM at inference. Use htop to monitor on Pi 5.

---

🚀 Deployment

Push to Ollama Cloud (for team sharing)

ollama login  # Use your jettanakorn account
ollama push jettanakorn/JFOX-AI

Auto-start on Pi 5 Boot

sudo cp deploy/jfox-ai-agent.service /etc/systemd/system/
sudo systemctl enable --now jfox-ai-agent

---

⚠️ Safety & Limitations

• Not a flight controller: JFOX-AI provides advisory diagnostics only—never direct control without human oversight
  
• Garbage in, garbage out: Erroneous sensor data may lead to incorrect conclusions
  
• No real-time guarantee: Inference latency depends on Pi 5 load
  
• Always inspect physically: AI suggestions must be validated on the ground

🔒 Data Policy: All telemetry stays on-device. No data is sent to Ollama Cloud unless you explicitly push the model.

---

📚 References

• Ollama Modelfile Docs
  
• Baby Redbee Flight Controller (example link)
  
• MAVLink Rust Crate

---

🧑‍✈️ Developed By

Jettanakorn / JFOX Aircraft Co., Ltd.
Senior Aerospace Engineer & Flight Control Systems Lead

Part of the Readbee Team — Advancing autonomous VTOL drone technology.

---

🦊 JFOX-AI: Your silent co-pilot in the sky.
“When the data speaks, we listen.”

---

License: Internal Use — JFOX Aircraft Co., Ltd.
Version: v1.0 (2026-01-07)