About

Gruhesh Sri Sai Karthik Kurra

AI and machine learning engineer. I write the notes I wish existed when I was learning this stuff — full derivations, real numbers, and code that actually runs.

## What this blog is Every post here comes out of building something. When I worked through reverse-mode autograd, I wrote the autograd post. When I rebuilt GPT-2, I wrote that one. The pattern is always the same: derive the math by hand, write the naive version, break it, fix it, then check it against a reference implementation. That means the posts are longer than most. They keep the derivations in, they keep the parameter-count tables in, and they use the actual hyperparameters from the actual config rather than round numbers. If you want a five-minute summary, these are the wrong notes. If you want to be able to rebuild the thing, they should be enough. ## Background Currently doing an MSc in Artificial Intelligence at Imperial College London. Before that, working as an AI integration engineer — mostly LLM systems, retrieval, and the unglamorous plumbing that makes models useful in production. Interests, roughly in order: language model internals, training dynamics, numerical methods, and how much of "modern architecture" is genuinely necessary versus inherited. ## Where to start - **[AI from Scratch](/series/ai/)** — the whole thing in order, linear algebra through a working GPT-2. Start at part one. - **[Topics](/tags/)** — if you already know what you're looking for. - **[Library](/library/)** — books, courses, and papers I actually used, no affiliate links. - **[Papers](/papers/)** and **[News](/news/)** — what I'm reading, updated regularly. ## Elsewhere - Portfolio — [gruheshkurra.com](https://gruheshkurra.com) - GitHub — [GruheshKurra](https://github.com/GruheshKurra) - Hugging Face — [karthik-2905](https://huggingface.co/karthik-2905) - LinkedIn — [gruhesh-sri-sai-karthik-kurra](https://linkedin.com/in/gruhesh-sri-sai-karthik-kurra-178249227) - X — [@Karthik__kurra](https://x.com/Karthik__kurra) ## Get in touch Corrections are welcome and genuinely wanted — if a derivation is wrong or a number is off, tell me. Email [gruheshkurra2@gmail.com](mailto:gruheshkurra2@gmail.com). If a post saved you an afternoon, you can [buy me a coffee](https://buymeacoffee.com/gruheshkurra).