Automate EVERYTHING!

In the past couple of months, one would have to be living under a rock not to notice that the world just wants AI. What two years ago was Web3, today is just AI.

“You’re using your hands to write an email or draw an image? Disgusting! A true 2025 technophile never does anything by hand! We are not animals! Real 2025 technophiles ask for things and let the robots do their thing.”

Open LinkedIn and you’ll find only posts bragging about amazing pipelines people have automated and how they got rich after selling their vibey apps. Not only that, most of those posts are clearly written by AI. (My personal record is probably around five sentences before my eyes just give up on those emoji rich lists upon lists of explanations.) 

 I guess LinkedIn in 2025 is what Facebook was in 2015: instead of being a platform where we see promoted posts from our network or valuable articles, we’re stuck in the early stages of an algorithm desperately trying to keep us hooked. These days, my LinkedIn feed looks more like my YouTube feed, which isn’t exactly useful for networking, finding talent, and so on. And the all-powerful algorithm sees only one thing: AI.

Don’t get me wrong, I love AI and would consider myself quite the “power user.” I’ve released a couple of mobile and web applications with a somewhat level of complexity and features. Not to mention all the cool things my team and I built at PlayPack regarding image, video and story AI driven pipelines for games.  What I see in the market of “AUTOMATE EVERYTHING” feels in most cases like a panic reaction to a new technology. Obviously it’s most likely the most disruptive technology we are seeing since the .com boom and the market will need a couple of years until we all understand what’s really possible and what not. 

My take on automation really is very straight forward: do you have a problem that needs automation? If the problem is complex enough and repetitive enough that having a real person do it would literally be a waste of time then is clearly a yes. But if you think that an AI will be able to create a better user experience than a passionate customer care specialist or scale a complex product better than a seasoned senior developer or to have a higher company impact than even a very passionate, curious and hard working junior developer, you are in for a surprise. 

Since I’m starting to look into new opportunities, and the only thing the market seems to demand is AI-savvy profiles, I’ve decided to go for the “show it, don’t tell it” approach. 

Instead of explaining in my cover letter how at my previous company we built AI pipelines that outperformed teams of 50 artists from top-tier companies with just two amazing in-house artists, or how I handled customer care, while still  doing my product, team management and board responsibilities, for a game with over 800k DAU in 10 different languages, I decided to just create an AI avatar version of myself to do that for me.

I created a bunch of context about my CV, career, and projects. Wrote detailed context for each of the projects, companies, challenges, learnings and created a script to ‘chunk it’ and feed it into an LLM for context and voila! built a clunky virtual version of me who can somewhat reliably talk to anyone who might be interested in my career. 

Why talk with recruiter bots when you can have your own virtual AI avatar do the heavy lifting for you?

Irony aside, I must admit this was an interesting project. In the past, I’ve integrated LLMs into hobby projects to automate macronutrient breakdown and calorie counting from a photo or to schedule workout routines based on user profiles and needs. These felt very straightforward implementations and usage  of the LLM APIs without much need for context or complex rules. 
The ‘AI virtual me’, though,  was much more challenging. Making sure hallucinations were minimal, preventing it from breaking character when pushed, or ensuring it was “context aware” was way more harder than they looked from the outside, at least for my puny brain.

Building this ‘AI virtual me’ has certainly taught me about the importance of  good and dedicated QA hours to find all the many ways it can break or hallucinate. I’ve learned  that you need many datasets with very clear hierarchies and rules balanced at different strengths to get somewhat decent results. 

This “virtual avatar” might be no more than a silly gimmick but building one is something that I can only recommend to anyone 

I should probably spend a few weeks creating more context and feeding it more data chunks for my “virtual self,” but time’s up. ( We all know how these projects are addictive and never truly finished.)

Now I need to decide whether to redo it with a vector database instead of a traditional one, just to learn the differences between these firsthand or maybe I’ll just build some more Duplos with my daughter instead.

In a time when you can literally build anything, there’s just never enough time in a day to do it all.

I guess I’ll find out soon enough.