Engineering Notes
"The best systems are boring."
Welcome to my personal knowledge base. This site documents production-grade patterns, trade-offs, and system designs for building scalable software and AI systems.
The goal is to move beyond surface-level tutorials and codify the "Context-as-Code" required to steer autonomous agents.
🧠 Agentic AI & SDLC
Standardizing how we build software with AI agents.
- AI SDLC: A framework for Context-as-Code (
.cursor/rules, Skills, PRDs) and MCP integration to build consistent, high-quality software with agents.
🔬 LLM Foundations & Research
Deep dives into transformer architectures, training, and alignment techniques.
- Transformer Architecture: From attention mechanisms to GPT - understanding self-attention, positional encodings, and modern variants (RoPE).
- LLM Pre-training: CLM vs MLM objectives, training dynamics, warmup schedules, and scaling laws.
- Alignment: RLHF vs DPO: How LLMs learn to be helpful - reward modeling, PPO, and direct preference optimization.
- Parameter-Efficient Fine-Tuning: LoRA, quantization (4-bit/8-bit), and QLoRA for efficient fine-tuning.
- Long Context LLMs: Memory vs retrieval trade-offs, hybrid approaches, and the "lost in the middle" problem.
- LLM Evaluation: Beyond BLEU scores - automated metrics, LLM-as-judge, and production patterns.
⚙️ Machine Learning Systems
Productionalizing ML models is harder than training them.
- MLflow on Cloud Run: Serverless architecture for ML metadata and model registry.
- ONNX Java Serving: High-performance CPU & GPU inference in Java environments.
🛠️ Developer Experience
Tools and patterns for efficiency.
- Cloud Cheatsheet: CLI references for AWS/GCP/Azure.
- Zsh Setup: A blazingly fast terminal environment (Powerlevel10k, SDKMAN).
About Me
Maintained by Dheeraj Joshi, a Staff Systems & Machine Learning Engineer focused on large-scale personalization and agentic AI systems.