Neural Networks, Transformers, and How LLMs Learn
The machinery the deep dive skipped: how a neural network and the Transformer work, what loss really means, and how reinforcement learning trains a model.
Hi, I'm Abhimanyu.
Writing ahead of certainty — notes on the things I'm working to understand.
In databases, a write-ahead log records changes before they're applied — so nothing is lost if something goes wrong. This blog works the same way. I'm a software engineer who writes things down before I'm sure I have them right: deep dives into databases, profiling, the layers most code never touches, and the occasional note-to-self that turned into something worth sharing.
Start here
The Crypto Module: Hashing, Encryption, and What to Use When
A practical guide to cryptography primitives available in every language — hashing, salting, HMAC, AES encryption, and secure random. Know what to use when.
Series
All series → Database Internals 8 parts
1NF through 5NF Explained · What to Use When · How Databases Actually Store and Find Your Data · From Binary Trees to Database Indexes · Internals, Types, and Trade-offs · Reading and Understanding Query Plans · Strategies, Real Advantages, and Pitfalls · Connections, Locks, and MVCC
How LLMs Work 3 parts
Intro to Large Language Models · A Deep Dive into How LLMs Are Built · Neural Networks, Transformers, and How LLMs Learn
Recent posts
All posts →-
-
Stop Re-Explaining Yourself to Your AI Coding Tools
Make Claude, Codex, and Gemini efficient: externalize knowledge into files, turn repetition into commands, skills, and hooks, and stop starting from scratch.
-
Postgres Concurrency Internals: Connections, Locks, and MVCC
A practical guide to how Postgres handles many concurrent clients — process-per-connection, pooling, lock manager, MVCC tuples, deadlocks, and pg_locks.
-
Database Replication: Strategies, Topologies, and What to Use When
A practical guide to PostgreSQL replication — physical vs logical, sync vs async, single-leader / multi-leader / leaderless, and replica-read rules.