I’ve been fascinated by AI/ML for a while now. Not the hype around it, more the actual math underneath, the idea that you can teach a machine to learn a pattern instead of hardcoding every rule yourself. It’s always felt a little like magic to me. The problem is I never really knew where to start.
I’ve built stuff before: quant trading systems, finance terminals, risk models — so it would’ve been easy to just lean on that and bolt a pre-trained model onto some project to say I “did AI.” I decided not to do that this time. I want to actually understand what’s going on underneath, from scratch, like I know nothing. No shortcuts, no importing a library and calling it a day.
So instead of grinding through tutorials in a vacuum, I picked something real to aim at: kaggriculture, a farming-simulation competition. Having an actual leaderboard and messy real data in front of you does something tutorials just can’t, you learn why your model is under-fitting because your own score is sitting near the bottom, not because a textbook mentioned it in passing.
The competition placement is really a secondary goal though. What I actually care about is getting the fundamentals right this time:
- what’s actually happening when a model “learns”: gradient descent, loss functions, the optimisation itself
- how much of your ceiling gets decided by data quality before you’ve even touched a model
- why boring stuff like train/val/test splits and avoiding data leakage matters more than picking a fancier architecture
- learning to read why a model fails, not just look at the final score and move on