Wanna build a real world LLM&ML system with me?

I am no AI influencer.

I am an ML engineer who builds ML/LLM systems for a living.

I have been doing that for 9 years.

And do you know what I found out?

Companies (still in 2025!!) struggle to find ML engineers who can design, implement and bring to life ML and LLM systems that move business metrics.

This is the ML expertise that I share in this newsletter, and in each of my courses.

Next Monday I will start again teaching my live course, Building a Real Time ML System. Together. A 15-session (50 hours) live coding experience in which I show you hands-on all I have learned this past 9 years: a universal pattern and a bag of tricks to build ML systems that react fast to data.

On the way, I will also show you how to fine-tune and deploy LLM apps that work in a real world business setting.

Let me quickly tell you what’s the plan.

The plan

I teach you how to build cars. Like real cars, not toy cars.

In this case, we will build a real time ML system to predict short-term crypto prices.

I will first show you how to setup a local Kubernetes cluster with Docker (with the help of Marius), so we develop production-ready code from day 1.

The first pipeline of our system will ingest live trades from Kraken, and will produce (hopefully useful) ML model features using technical analysis.

After that, I will show you how to plot this data in real-time with Grafana.

Then we will enter the ML modelling phase, and we will write the training pipeline of our system. We will deploy it to our Kubernetes cluster and implement automatic re-training.

Once we have a model, I will show you how to produce and serve its predictions in real-time. We will implement this component in Rust, because I think this is the language you should use to write this part of the system (you will hate me a little bit for not using FastAPI here, but I promise you the effort is super worth it).

Once we have this end-2-end raw-data-to-predictions system up and running, we will try to improve it.

How?

We will add a second source of data (in this case financial news in real-time) and transform in real-time into a crypto sentiment score, with the help of an LLM.

I don’t want us to build a simple wrapper around a third-party LLM API, like OpenAI, Claude or Grok.

I want us to do some real world LLM engineering, so if you zoom in into this component of the system, you will find yet-another ML system :-) (this is Compound AI, my friend), with

• A feature pipeline that generates (instruction, output) datasets.

• A fine-tuning pipeline that produces a light 3B-7B LLM that works well and fast

• An inference pipeline using vLLM to serve the final model predictions to the rest of the infrastructure.

Look, I am not going to lie to you.

It will be tough.

And you will need to work A LOT.

But let me tell you something:

If you manage to do it, you are on the top 5%.

Do you accept the challenge?

Gift 🎁

As a subscriber to the Real World ML Newsletter you have exclusive access to a 40% discount. I need to pay a mortgage, and occasionally go on holidays, so grab it before I change my mind. 😉

See you on the other side,

Enjoy the weekend,

Pau