Five years ago I was happily buried in Selenium scripts, flaky test failures, and endless API assertions. Today, I’m building something very different: AI agents that think, retrieve knowledge, validate data, and take actions.

This post is a snapshot of how I moved from traditional QA into the world of Agentic AI - not by watching courses, but by shipping real systems.



Where I started

I have 8+ years of experience in test automation and quality engineering. My job was to make systems reliable:
finding edge cases, breaking APIs, and forcing software to behave correctly.

That mindset turned out to be my biggest advantage when I entered AI.

Most people treat LLMs like magic.
I treated them like unreliable APIs that need validation, constraints, and control.



The moment everything clicked

The shift happened when I built my first AI-powered dental appointment booking agent.

Not a chatbot demo.
A real system that:

  • Asks users for missing info
  • Validates dates, times, and phone numbers
  • Uses clinic knowledge via RAG
  • Confirms bookings
  • Handles corrections like “Actually make it Friday at 4”

I built it with PydanticAI, structured output schemas, and a retrieval system.

That’s when I understood:

AI doesn’t become useful when it talks better. It becomes useful when it can act safely inside constraints.



Why PydanticAI changed everything

Most GenAI projects fail because:

  • The output is unstructured
  • Nothing is validated
  • Agents hallucinate actions
  • You can’t trust the system

PydanticAI fixes that.

I use it to define:

  • Appointment schemas
  • User intent models
  • Tool inputs
  • Memory structures

The LLM is no longer free‑text chaos.
It’s a component in a real software system.



RAG, but inside the agent

Most people bolt RAG outside their app.

I learned something deeper:

RAG belongs inside the agent’s reasoning loop.

My agent doesn’t just fetch data. It uses retrieval as part of how it decides what to do.

Clinic hours, pricing, policies, services - all of it lives in vector search and is pulled only when needed.

This makes the agent:

  • More accurate
  • More controllable
  • Easier to update without retraining



QA → AI Engineer is not a leap. It’s a straight line.

Here’s the secret nobody tells you:

Good AI systems need the same things good test frameworks need:

  • Validation
  • Determinism
  • Observability
  • Reproducibility
  • Guardrails

That’s why my background in automation wasn’t a weakness - it was the bridge.

I don’t build prompt demos.
I build systems that fail safely.



What I’m building now

You’ll find on my GitHub:

  • Agentic AI systems with memory and retrieval
  • LLM‑driven API test generators
  • Pydantic‑validated workflows
  • Experiments with hindsight memory and self‑correction

Everything is designed around one idea:

If AI is going to run parts of the real world, it must be engineered like real software.


This journey is just getting started.
And this time, the code actually matters.