From Theory to Play

 My Journey into Agent RL and AI Tools

By Keerthana Venkateshkumar

Attending the recent online meetup led by Raj Marri, President at Cerebrone.ai, felt less like a lecture and more like stepping into a live AI playground. Concepts that often feel abstract agent reinforcement learning, RLHF, LoRA, and QLoRA were broken down, connected, and brought to life through hands-on demos. What stood out most was how seamlessly theory turned into practice.

Making Sense of Agents and RLHF

The session began with a simple yet powerful way of understanding reinforcement learning with human feedback (RLHF). An agent, as Raj explained, is not just a model that predicts outputs—it’s a system that learns through interaction. By taking actions, receiving rewards, and adjusting its policy (its behavior), the agent gradually improves.
What really helped was the explanation of critics—human or model-based—that guide the agent toward better decisions. This made it easier to grasp how modern AI agents can handle long, multi-step tasks, whether that’s navigating interfaces, calling APIs, or making sequential decisions in complex environments.

When Theory Meets Play: Tic-Tac-Toe

The highlight for me was watching an LLM agent learn to play Tic-Tac-Toe. Using the OpenPipe/ART notebook (it didn't work in my system though), we saw how an environment is formally defined by board states, legal moves, and rewards for wins, draws, or losses.

(THE GAME GOES LIKE...) The agent starts off clueless, playing random moves. But as it plays hundreds of simulated games, it collects trajectory records of actions and outcomes. These trajectories are then used to fine-tune the model. Slowly, the agent stops making silly mistakes and starts playing strategically. Seeing this learning curve unfold made reinforcement learning feel tangible and intuitive.

GRPO: Learning Through Comparison

Another concept that blew my mind out was Group Relative Policy Optimization (GRPO). Instead of evaluating one output at a time with expensive human labels, GRPO compares multiple outputs or trajectories against each other and learns from their relative quality.
This approach feels far more practical for real-world systems, where absolute “perfect” labels are rare. It showed how scalable agent training can be achieved without depending heavily on costly reward models.

LoRA and QLoRA: Making Fine-Tuning Practical

-- never heard of these or learned about them until i joined this meetup -- 

" LORA-  low rank adaptive fine tuning "

Fine-tuning large models often sounds intimidating because of hardware constraints. That’s where LoRA and QLoRA come in. LoRA updates only a small subset of parameters using low-rank matrices, while QLoRA goes a step further by running the base model in quantized (low-precision) form.
In the demo, these techniques made it possible to fine-tune a large model on a T4 GPU, proving that meaningful experimentation doesn’t always require massive compute resources. It was reassuring to see how accessible advanced fine-tuning can be with the right tools.

From Notebooks to  Workflows

What I appreciated most was the end-to-end perspective. This wasn’t just about concepts it was about doing the work. From setting up the environment and running notebooks to hosting models with vLLM, collecting rollouts, applying GRPO, and evaluating performance with clear metrics, the session provided a realistic blueprint for building agentic systems.
It clearly showed how modern AI tools, efficient fine-tuning methods, and reinforcement learning come together in practice.

Curiosity, Questions, and What’s Next

The open Q&A made the session even more valuable. Questions around when to use RL versus supervised learning, how to design reward signals, and how to move from demos to real projects were openly discussed.
For me, the meetup sparked ideas beyond games towards agents that can automate workflows, interact with software systems, and adapt to complex environments. I walked away not just with notes, but with motivation to experiment, build, and explore more agent-based AI systems.

i even got to personalize my learning experiences and what's and what nots to do in portfolio etc.

 ps : my first article!!!!