Daddy-Myth: A Full-Stack Practical Atlas for AI Systems Engineers

This article showcases an AI systems engineer's complete project practices in RAG systems, multimodal learning, deep learning, and other fields, covering end-to-end experience from underlying model implementation to industrial-grade deployment, reflecting end-to-end AI engineering capabilities.

In the AI field, there are many people who can train models, but relatively few who can fully deploy AI systems to production environments and ensure reliability. End-to-end engineering capabilities need to span technical stacks such as understanding underlying model architectures, training optimization, inference deployment, and system integration. Daddy-Myth's GitHub profile covers key areas like RAG retrieval systems, multimodal visual question answering, object detection, and reinforcement learning alignment, with each project considering actual deployment scenarios.

Core Capabilities

• Programming languages: Primarily Python, PyTorch, Transformers; also has system-level development skills in Java, C/C++
• LLM systems: RAG pipeline, retrieval optimization, embedding models, vector databases, RLHF alignment
• Computer vision: OpenCV, ViT, Faster R-CNN, full object detection workflow
• Security basics: Practical experience with Nmap, Wireshark, Metasploit

Engineering Philosophy

1. Reliability first: Focus on stability in real environments rather than just metrics
2. Zero-cloud awareness: Multiple projects support local deployment to reduce external dependencies
3. Modular design: Decoupled components for easy maintenance and expansion
4. Complete workflow: Covers the entire process from data processing and training to deployment interfaces

RAG Systems

• D_RAGon_Psyche: A professional retrieval system for psychology, using BGE-M3 embedding model, academic text chunking strategy, domain fine-tuning, and supporting zero-cloud local operation
• D_RAGon_System: A general modular RAG framework that provides complete components such as document ingestion and embedding generation

Multimodal Learning

• Visual Q&A System: An end-to-end visual question answering system using DINO ViT (visual encoder), DistilBERT (text encoder), GPT-2 (decoder), and a custom multimodal fusion architecture

Computer Vision

• Faster-RCNN-From-Scratch: Implementing Faster R-CNN from scratch, including core components like RPN and ROI pooling
• PCB-Defect-Detection-System: Industrial-grade defect detection using YOLOv8 for localization, classification, and severity assessment

Large Model Alignment & Fine-tuning

• Fine-tuning-Flan-T5-RLHF: Exploring the application of RLHF in text summarization tasks
• GPT-2 LaTeX Generation: Fine-tuning GPT-2 to convert natural language to LaTeX code

Other projects: AnimeGenreBERT (multi-label classification), FaceID Pipeline (face recognition workflow), WebStrike (web vulnerability scanner)

Daddy-Myth's GitHub profile is a growth example for AI systems engineers, showing how to gradually build complete capabilities from theory to practice and from experiment to production through in-depth projects. The reference value of this path lies in: not being satisfied with calling ready-made APIs but deeply understanding underlying mechanisms; not only focusing on model training but considering complete deployment scenarios; driving learning with actual projects and accumulating capabilities while solving real problems.

1. Depth over breadth: Underlying implementations (like developing Faster R-CNN from scratch) accumulate irreplaceable engineering judgment
2. Project-driven learning: Choose projects that solve real problems (such as psychology RAG, PCB defect detection) instead of mere exercises
3. End-to-end thinking: Understanding the full system from data preparation to model deployment is crucial for designing reliable production systems