Phase 1
Foundation
Completed Infrastructure- Home AI server configured
- Wildlife dataset prepared
- First custom bush turkey model trained
- Outdoor camera installed
- Garden Guard website deployed
- Power Apps MVP published
Log #02· Published October 2026
From AI Model
Training to a Working
Monitoring Platform
Garden Guard has progressed from training its first wildlife detection model to testing a complete monitoring workflow. The project now combines local AI processing, outdoor camera footage, cloud infrastructure and a Power Apps review dashboard.
01. The Hardware & Infrastructure Setup
I set up a Reolink camera feeding live video to my home AI server, where YOLO analyses the stream using an NVIDIA RTX 3060 GPU. The system now runs continuously through Docker, automatically restarts after a reboot, and deploys updates from GitHub through a self-hosted GitHub Actions runner.
services:
yolo-detector:
image: gardenguard/yolo-detector:latest
runtime: nvidia
gpus: all
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
restart: always
environment:
RTSP_STREAM: configured in environment
MODEL_PATH: /app/models/best.pt02. Detection Image Storage
When the camera detects an animal visit, Garden Guard saves the best frame locally on the Linux server. Images are automatically organised by date and detected species, making them easier to review and reuse for training.
03. Training the Custom Bush Turkey Detector
I also started building a custom bush turkey detector. I collected and annotated 55 bush turkey images in Roboflow, exported the dataset, and trained my first YOLO model locally on the RTX 3060 using PyTorch. The first training run completed successfully and produced a custom best.pt model. Its initial validation results were promising, but the dataset is still small, so this is an early prototype rather than a reliable final model.
04. Deployment
Deploying the Garden Guard Platform
The Garden Guard website is now deployed as a live cloud-hosted platform. Its deployment workflow uses containerisation, automated releases and secure web infrastructure to make future updates faster and more reliable.
CI/CD Pipeline & Architecture Map
Live in production- GitHub
- GitHub Actions
- AWS Lightsail
- Docker Compose
- Caddy
- gardenguard.org
Web Platform Stack
- Frontend Framework
- Next.js & Tailwind CSS
- Host OS
- AWS Lightsail Ubuntu
- Containerization
- Docker Compose
- Reverse Proxy
- Caddy / with SSL
- DNS Management
- Amazon Route 53
- Transport Security
- HTTPS / TLS 1.3
- CI/CD Pipeline
- GitHub Actions
05. Companion application
Building the First Detection Review Dashboard
A Microsoft Power Apps MVP has been created and published to test how users could review and classify Garden Guard detections before the custom web dashboard is developed.
01 / Dashboard
Total visits1,284Confirmed94.2%
Recent Activity Log• Bush Turkey — Cam 01 (06:42)• Brush-tailed Possum — Cam 02 (03:14)
02 / Detection review
Predicted: Australian Bush TurkeyConfidence Score: 94.2% · Camera 103 / Save review
✓Review Saved SuccessfullyThe classification has been logged into the SharePoint database.
06. Data pipeline
Designing the Detection Data Pipeline
SharePoint has been selected as the initial metadata and workflow layer for the MVP. It will connect the Linux-based AI system with Power Apps and provide structured records for reviewing wildlife detections.
Pipeline Integration Layout
- 01Outdoor CameraLive video capture of the backyard garden
- 02Ubuntu AI ServerDockerized stack running continuous inference
- 03YOLO Detection & MetadataConfidence, species, and image timestamp output
- 04SharePoint ListStructured workflow repository with webhook triggers
- 05Power Apps Review DashboardSecure review interface for classifying visit events
- 06Reviewed Detection RecordFinal archived conservation dataset
07. Project horizons
The Road Ahead
Our technical roadmap maps out the research horizons of the Garden Guard platform. Rather than rapid scaling, our milestones prioritize deeper backyard insights, open-source documentation, and sensible local automation.
Phase 2
Automation
In Progress Development- Connect live camera recordings to the AI pipeline
- Automate metadata transfer from Linux to SharePoint
- Replace Power Apps sample data with real detections
- Improve wildlife model accuracy
- Add notification and nightly-summary workflows
Phase 3
Smart Garden
Future Research Goals- Custom web-based monitoring dashboard
- Soil-moisture monitoring integrations
- Rainfall and environmental data charts
- Wildlife behaviour and pattern analysis
- Humane deterrence and garden-protection experiments
Current Project Status & Progress Tracker
What's Working
- AI server and GPU environment
- Labelled wildlife dataset
- First bush turkey detection model
- Outdoor camera recording
- Production website and CI/CD
- Power Apps MVP interface
What's in Development
- Automated YOLO inference pipeline
- Linux-to-SharePoint integration
- Real detection data in Power Apps
- Improved model accuracy
- Nightly monitoring summaries
Future Research
- Soil and rainfall sensors
- Custom web dashboard
- Wildlife behaviour insights
- Humane deterrence methods
- Smart garden recommendations
“Garden Guard is no longer only a model-training experiment.
It is becoming a complete system for observing, understanding
and responsibly responding to backyard wildlife.”