One skill every ML engineer has to master ↓

𝗠𝗟 𝗦𝘆𝘀𝘁𝗲𝗺 𝗱𝗲𝘀𝗶𝗴𝗻

Yes. And do you know why?

Because good ML system design has NOT changed at all in the last 5 years.

And it won't.

𝗪𝗵𝘆 ❓

Because any ML system is (and will always be) made of 3 types of programs (aka pipelines)

1️⃣ → 𝗙𝗲𝗮𝘁𝘂𝗿𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀 that transforms raw data into ML model features (e.g. vector embeddings) that are saved in a Feature Store or Vector DB.

2️⃣→ 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴/𝗙𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀 that read historical features from the Feature Store/Vector DB and generate a new model artifact, either by training from scratch or fine-tuning a base LLM. This model artifact is then pushed to a model registry.

3️⃣ → 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀 that load the model from the registry, and the input from the client app (for example a vector of numerical features, or a text prompt), generate a prediction (or a generation) and return it to the client app.

This is a universal blueprint, that together with CI/CD workflows (aka MLOps) helpx you build any ML system.

Now you can go and thank Jim Dowling for the idea 💡

----

Hi there! It's Pau Labarta Bajo 👋

Every day I share free, hands-on content, on production-grade ML, to help you build real-world ML products.

𝗙𝗼𝗹𝗹𝗼𝘄 𝗺𝗲 on Substack so you don't miss what's coming next

#mlops #machinelearning