A school for people who build with AI

What is here

The courses are the centre of it: each takes one subject and works through it in order, a part at a time, often with a companion video. The blog carries standalone essays on what broke in production and why. The papers section explains research worth knowing, and the interview notes are revision material filed by topic.

Who writes it

I’m Rajan Devkota, a Senior AI Engineer based in Pokhara, Nepal, and I build and operate production AI systems for a living.

Most of my work sits at the seam between a language model and the documents it is supposed to reason over. That seam is where projects actually succeed or fail: how a page is segmented, what survives transcription, which retrieval strategy is used, and whether anyone notices when it quietly degrades.

Lately that has meant a two-stage document pipeline that separates layout detection from transcription, long-horizon extraction agents that stay coherent across hundreds of documents, and a temporal memory layer on Postgres and Qdrant whose in-memory index is 32× smaller after binary quantisation — with recall held up by rescoring rather than hope.

A good deal of it is unglamorous cost engineering: moving a serving path onto cheaper GPUs without losing throughput, autoscaling inference so it is not paid for while idle. The recurring lesson is that the interesting engineering is almost never in the model.

Why it exists

Writing is how I make sure I actually understood something, and teaching it in order is the harder test. The school is where that gets published, in case it saves someone else the same months. The systems behind the lessons are written up under projects.

If you are a researcher, a hiring manager, or someone who has run into the same problems — I would genuinely like to hear from you.