Harvard’s John A. Paulson School of Engineering and Applied Sciences (SEAS) wanted to offer a serious, practitioner-level program on large language models. Not a high-level overview for executives, but a hands-on course where CTOs, consultants and technology leaders could actually build things. They needed someone who had been in the trenches building AI products, not just studying them.
The challenge
The field was moving fast. RAG, agents, reasoning models, fine-tuning. There was no shortage of blog posts and tutorials, but most of them either stayed too theoretical or relied heavily on abstraction layers that hid what was actually going on (Langchain anyone?).
Harvard’s teaching faculty wanted course material that cut through that: practical, grounded in real experience, and deep enough for a technical audience that would push back on hand-waving.
What we did
We built the entire course from scratch. Slides, hands-on coding exercises, and live demonstrations covering the full arc from transformer basics to agentic AI. The curriculum spanned:
- Introduction to language models and deep neural networks
- Transformer architecture (GPT)
- Instruction-based models: from GPT to ChatGPT
- Retrieval-Augmented Generation (RAG), Self-RAG, and Graph-RAGs
- Fine-tuning LLMs using LoRA
- AI agents
The approach was deliberately first-principles. For RAG, instead of jumping to off-the-shelf libraries, we broke it down to a simple demonstration that made the mechanism click. For reasoning models like DeepSeek, we showed how they’re fundamentally different from assistant models: they aren’t trained on ideal responses, but improve by learning from their own outputs in verifiable domains like math and coding. That distinction matters when you’re deciding what to deploy in production.
The program was delivered in person to 300+ participants across a multi-day format, with guided coding sessions and an optional hackathon capstone.
The result
The course is now Harvard SEAS’s official “Large Language Models: From Transformer Basics to Agentic AI” program, running as a recurring live online course. The material we built, the slides, exercises and demonstrations, is what Harvard’s faculty uses to deliver it on an ongoing basis.
As Pavlos Protopapas, the Program Director, put it: “The hands-on skills and foundational understanding you’ll gain in this course will make you a more effective AI professional.”
