Overview

Gen-AI Engineering Series

Gen-AI Engineering Series
saraththarayil
saraththarayil
· February 5, 2026 · 4 min read

From Scattered Knowledge to Systematic Understanding

You know that feeling when you’ve picked up skills here and there? A tutorial on RAG systems one weekend, some prompt engineering during a hackathon, deploying a model for that one project. But you’ve never actually sat down and learned everything properly from start to finish?

That’s me with Generative AI.

I can build things. I’ve shipped AI features to production. But there are gaps. The kind where someone asks “how exactly does attention work?” and I mumble something about matrices and change the subject. Or when I’m debugging a RAG pipeline and I’m not sure if my chunking strategy is brilliant or backwards.

This series is my attempt to stop being a GenAI magpie (collecting shiny techniques without understanding the full picture) and actually travel a systematic learning journey from foundations to production systems.

Note

This isn’t a beginner’s guide. It’s a practitioner’s refresh. Going back through the fundamentals I thought I knew, filling in the gaps I didn’t know existed, and building a complete mental model of how modern GenAI systems actually work.

The Problem

What I already know (sort of):

  • Writing effective prompts… most of the time
  • Built a RAG system… that worked… eventually
  • Understand embeddings… conceptually… ish
  • Deployed models… with varying success
  • Agents are a thing… but I’m fuzzy on details

What I’m tired of:

  • Googling the same concepts repeatedly
  • Copy-pasting solutions without understanding why
  • Surface-level knowledge of everything, deep understanding of nothing
  • Building things that work but can’t explain how
  • Starting every project from scratch

This sprint converts scattered tactical knowledge into strategic understanding.

đź“š Sprint Structure

Building knowledge systematically, each concept on top of the last. What I should have done the first time.

1. LLM Foundations & Prompt Engineering

Note

How LLMs actually work: Understanding attention, tokenization, and design choices. Not just “transformer architecture” as a black box.

  • Prompt engineering: Systematic techniques. Few-shot learning, chain-of-thought, when each helps
  • Token mechanics: Why context windows matter, how tokens are counted, why costs vary 10x
  • Model selection: When to use which model instead of defaulting to GPT-4

Fixing: Treating prompts like incantations and hoping they work

2. Retrieval-Augmented Generation (RAG)

Summary

Vector embeddings: What they represent, why models matter, how similarity search works. Not just “convert text to vectors.”

  • Chunking strategies: Understanding trade-offs instead of trying sizes until something works
  • Retrieval pipelines: Engineering quality. Precision vs recall, different methods, debugging
  • Evaluation: Proper metrics, testing, iteration beyond “seems fine”
  • Advanced patterns: Hybrid search, re-ranking, query transformation

Fixing: Throwing more documents at the problem when retrieval is poor

3. Azure AI Foundry & Deployment

Summary

Infrastructure as code: No more clicking through Azure Portal at midnight

  • Monitoring: Understanding what’s happening, why things break, catching problems early
  • Cost management: Where money goes, how to optimize without breaking things
  • Scaling patterns: When to scale, how to scale, when you’re over-engineering
  • Security: The stuff I skip but shouldn’t

Fixing: “Deploy and pray” approach

4. AI Agents & Model Context Protocol (MCP)

Summary

Agent architectures: ReAct, plan-and-execute, when to use which pattern

  • Tool calling: How agents use tools, designing good interfaces
  • MCP: New standard for interoperable AI systems
  • Error handling: Agents fail in creative ways
  • Multi-agent systems: When they make sense vs over-complicating

Fixing: Vague hand-waving about “agentic workflows”

5. Capstone Project

Summary

A real project that:

  • Solves an actual problem
  • Uses multiple sprint components
  • Is deployed and accessible
  • Has proper monitoring
  • I’d show to employers/clients

🗂️ Posts in This Series

đź”— Why Document This?

Writing forces clarity, and future-me needs documentation. Public accountability keeps me honest about finishing. If you have scattered knowledge too, this might help. If you’re an expert, tell me where I’m wrong. This is a systematic GenAI journey, honest about learning, focused on building working systems and documenting mistakes. It’s not a beginner guide, academic survey, or expertise flex. Just learning in public.

🚀 Let’s Go

The goal is systematic understanding building a clear mental model instead of collecting disconnected tricks. If you’re also learning GenAI seriously, feel free to follow along, try the projects, and share feedback. This is learning in public.

P.S. I’ll use AI tools throughout this journey, transparently and thoughtfully.