Patents Are Not the Main Work

Last week I got two emails from the patent team: two more of my patent applications are going to issue as US patents.

This will bring me to 7 issued patents, with roughly a dozen more still pending review.

The two new ones are around document understanding and LLM explainability. I am intentionally not putting the full titles here, because the titles are long and, honestly, patents are not the main point.

A patent sounds like a clean invention story.

In practice, at least in applied AI, it is usually messier than that. You run into a specific production problem, try a few things that do not work, combine some existing ideas in a way that fits the domain, and only later realise there may be something novel enough to protect.

That part is nice, of course.

But patenting ideas is far from the main work of an AI engineer.

The main work is less shiny:

• choosing the right problem
• talking to users before building too much
• killing the first idea when it is wrong
• translating ML metrics into business outcomes
• making the system reliable enough to run without you watching it
• keeping costs under control after the demo works

On June 15, I’ll be joining DataTalks.Club live to talk about this side of AI engineering: how to build AI that actually ships in production.

We’ll talk about the agentic system I worked on at Intuit, why the original chatbot idea did not survive, how evaluation and business metrics shaped the project, and why the model is rarely the win.

Join live if you want to ask questions in real time:

https://luma.com/6c089dac

P.S. Patents are not a default part of AI/ML/DS work. A lot of excellent people in the field never write one. I was lucky to learn how the process works a few years ago from very good teachers, and to work in places where turning practical ideas into patent applications was supported. It is a separate skill, and I am still grateful someone taught me.