Learning
Deep-dive writeups from ongoing research projects, grouped by topic. Each one ends with an optional multiple-choice quiz on the core concepts.
Retrieval Research
Why an LLM agent misses the right doc, how far you get without vectors, and whether a small fine-tuned retriever beats it.
Teaching a Small Model to Find Documents
Bi-encoders, cross-encoders, retrieve-then-rerank, and contrastive fine-tuning — a field guide to how a small model learns to match queries to documents.
Inside the Transformer
From raw text to the 384 numbers a bi-encoder hands back: tokenization, attention, pooling, and why training one from scratch on a small corpus wouldn't work.
Cosine Similarity
The geometric intuition for why direction, not distance, is what retrieval ranks documents by — with an interactive demo you can drag around and a worked ranking example.
The Retrieval Pipeline, Line by Line
Three files that turn bi-encoder architecture diagrams into something that actually runs: synthetic training pairs, contrastive training, and brute-force cosine search — with the real code for each.
The Best Wrong Answer
A hard negative is a document that looks right and isn't — and it matters more to a contrastive loss than almost anything else in the batch. The math for why, worked through step by step.
Digging for Hard Negatives
The code that mines real hard negatives per base model, a real CUDA out-of-memory crash mid-sweep, and a 12-run sweep in which no configuration clearly beat its in-batch-only baseline.