I hope you enjoyed experimenting with different LLMs locally on your machine using tools like Olama. In my own experiments I found some interesting differences: for example, Quinn 2.5 seems especially strong at understanding many languages, while other models sometimes explain concepts more clearly. I’d love to hear about your findings — post them in the course or message me directly. Finding the right model for your problem is a critical skill for any LLM engineer, and time spent exploring models is time well spent.
What you’ll achieve in eight weeks
Over the next eight weeks I’ll take you from where you are now to being confident in LLM engineering. Here’s the plan.
Week 1 – Frontier models
We’ll explore frontier, closed-source models (the latest previews and high-performing models) through web UIs like ChatGPT and via API access. You’ll build a compact commercial project that is immediately useful and gives a real-world exercise in integrating frontier models.
Week 2 – Add a user interface with Gradio
We’ll build a polished UI using Gradio. It’s perfect for quickly creating a sharp front end even if front-end work isn’t your strength. The plan: create a multimodal AI assistant with audio, images and tool use (the assistant will call out to code running on your machine).
Week 3 – Open source on Hugging Face
We’ll move to open source using Hugging Face tools. You’ll learn the Pipelines API for quick experiments and investigate tokenizers and models in more depth using the advanced APIs.
Week 4 – Model selection and benchmarking
With so many models available, choosing the right one is nontrivial. We’ll learn how to benchmark models, use leaderboards, and define a decision path to pick the best model for a given task. The hands-on challenge: build an application to translate Python to high-performance C++ and compare model outputs — one model will stand out with dramatic speed improvements.
Week 5 – Retrieval-augmented generation (RAG)
RAG is a hot topic. You’ll build a RAG pipeline that answers organization-specific questions from indexed documents. There’s a commercial challenge to apply this approach to your own data.
Week 6 – Start the flagship project
We’ll define a business problem, prepare data, and create traditional machine-learning baselines. Then we’ll experiment with frontier models and fine-tune them to push performance further.
Week 7 – Open-source fine-tuning
Open-source models may initially lag behind frontier models. We’ll fine-tune them aggressively with the goal of competing with GPT-4–class performance. Expect surprising results.
Week 8 – Finale: multi-agent autonomous solution
In the final week we’ll assemble everything into a fully autonomous agentic solution made of multiple collaborating agents that solve a real commercial problem, scan the web where needed, and even push notifications for discoveries. This will be the culmination of the prior weeks and produce a deployable solution you can use at work.
Wrap-up
Each week builds on the previous one — practical projects, model selection, fine tuning, and real commercial outcomes. If you’re eager to learn by doing and want to produce projects you can use right away, this eight-week path will get you there. Share your experiments, ask questions, and bring your curiosity — that’s what makes an LLM engineer great.