[D] Write simple "Case Study" article about experiences to integrate onnx on Android

[michalharakal]

Documentation Task: Case Study – Inference with YOLO runtime on ONNX vs SKaiNET

NOTE: Following the DARC develpmnet workflow we create documentation leaning on Diátaxis (Divio) documentation system.

Goal

Create an Explanation-style documentation page that describes a real-world journey of completing the same computer-vision task using two approaches:

• A traditional ML workflow (YOLO + ONNX)
• The SKaiNET platform

The purpose is to help readers understand the practical differences in workflow, complexity, focus, and tradeoffs between both approaches.

---

Scope

This document is not a tutorial or a how-to guide.
It will be a case study / experiential explanation focused on:

• How the same task was approached with both frameworks
• Where time and effort were spent
• Friction points and advantages
• What changed using SKaiNET
• Key lessons learned and decision guidance

No step-by-step instructions are needed beyond high-level descriptions.

---

Proposed Structure

1. Context

Describe the problem being solved (e.g., object detection with YOLO) and why both approaches were tested.

---

2. Traditional ML Journey (YOLO + ONNX)

Topics to cover:

• Environment setup and dependency management
• Model configuration and customization
• Dataset preparation and augmentation
• Training loops and hyperparameter tuning
• Model export to ONNX
• Inference setup and performance validation
• Experiment tracking and manual debugging

Focus on:

• Engineering overhead
• Tool fragmentation
• Complexity
• Time investment

---

3. SKaiNET Journey

Topics to cover:

• Dataset ingestion into SKaiNET
• Declarative training setup
• Platform-managed training and tuning
• Built-in logging and tracking
• Automatic scaling and optimization

Focus on:

• Reduced boilerplate
• Simpler mental model
• Shift in user effort from infrastructure to evaluation

---

4. Side-by-Side Workflow Comparison

Include a comparison like:

Area	YOLO + ONNX	SKaiNET
Setup complexity	High	Low
Code required	Extensive	Minimal
Training loop	User-managed	Platform-managed
Hyperparameter tuning	Manual	Automated
Experiment tracking	External tools	Built-in
Inference pipeline	Manually assembled	Integrated
Scaling	Custom engineering	Automatic
Primary effort	Infrastructure	Dataset & evaluation

---

5. Observations and Learnings

Summarize key takeaways:

• How the experience differed
• Where SKaiNET saved time
• Areas where traditional ML provided more control
• Surprises or unexpected challenges

---

6. Choosing an Approach

Provide guidance for readers:

• When YOLO + ONNX is preferable (research, fine-grained control, custom architectures)
• When SKaiNET is preferable (rapid development, managed pipelines, production workflows)

---

Deliverable

A single documentation file in Markdown or asciidoc format titled stored in docs folder:

“Training the Same Object Detection Model with YOLO + ONNX and SKaiNET: A Comparative Case Study”

---

Audience

• ML practitioners evaluating workflow tools
• Teams deciding between traditional ML stacks and integrated platforms
• Engineers new to SKaiNET seeking practical context

---

Success Criteria

The document should:

• Clearly communicate the experience differences between workflows
• Explain tradeoffs without bias
• Avoid instructional depth
• Serve as an educational comparison rather than a setup guide

[michalharakal]

There is a cool article about intergarting TFlite by Kashif Mehmood

https://proandroiddev.com/from-5-fps-to-30-fps-what-the-tflite-defaults-are-hiding-a8fbf0d20ff6