Journal

thinking out loud — The Blog

Notes on machine learning, retrieval systems, MLOps and building software in public. Written while I figure it out, not after.

Four years of building things I wasn't ready for

On starting projects slightly past my skill level, contests that ended in third place, and why the half-finished repos still counted.

Building a RAG pipeline taught me that retrieval quality beats prompt cleverness. Chunking is where most of the pain lives.

Same code, same data, different results. Docker didn't make ML magical — it made the environment stop lying to me.

I kept drawing straight lines for agent workflows. The useful ones were loops. Notes on state, nodes, edges, and when a graph is actually worth it.

I fought the framework for a week, then realized I wanted adapters, not magic. What I actually use from LangChain and what I ignore.

One is a sport. One is a drill book. Mixing them up made me worse at both until I picked a job for each.

CP is weird training — fake constraints, real pressure. What stuck after the rating graph stopped being the point.