It's tempting to skip the math and stack frameworks. I tried that. My agents were mysterious black boxes — when they failed, I could only guess why.
Studying probability gave me language for what I was seeing: priors, likelihoods, expected values, the difference between noise and signal in eval results.
Where it shows up daily
Eval design: is my agent actually better, or did I get lucky on five test cases? Basic statistics answers that honestly.
Retrieval and ranking: understanding scoring as probability makes RAG behavior predictable instead of magical.
Decision thresholds: when should the bot answer vs. escalate? That's a cost-benefit calculation dressed as a product decision.
My study approach
One concept, one implementation. I code every distribution and estimator from scratch in Python before using a library — slow, but it sticks.
Math first, models second. The models change monthly; the math doesn't.
Enjoyed this? Let's talk about it.
I write to think clearly. If this sparked an idea — a project, a disagreement, a collaboration — I want to hear it.
Get in touch