I’m Bhavish. I build systems and machine learning software from scratch, and keep it honest with an oracle and a benchmark. Alongside the projects below, I’ve reimplemented micrograd, minbpe, nanoGPT and GPT-2 from scratch, and I read llama.cpp and SGLang for what only production code teaches.
Projects
- AGE Aris
- A local project-management app for teams where some of the workers are AI agents. The task board lives in git, with one Markdown file per task.
- AGEIS
- A local-first intelligence engine in C++20. Corpus in, entity graph out, and a loop that hunts the intrinsic connections between entities.
- AGEION
- A data engine that finds alpha in the internet haystack.
- RSNA
- Knee-MRI abnormality detection under severe label scarcity.
- Model Gym
- A code-first course that builds a Llama-style language model stack from first principles, out to preference RL.
- snapstr
- A byte-level BPE tokeniser on its way to C++, where every optimisation has to beat a predicted number.
- wLLM
- A continuous-batching LLM inference server, native to Windows.
- LLM Release Gate
- A quality and safety framework for LLM apps, built in four hours at a hackathon. It derives tests from risks and recommends whether to release, with a person making the final call.
Writing
- The headline on this site is a tokeniser
How the homepage trains a byte-level BPE vocabulary in your browser, using the page's own text as the corpus.