monolog

About

updated 2026-01-01

I'm Avishake Adhikary — a machine learning engineer who builds AI from the ground up. Not wrappers around APIs, not fine-tuned adapters. I mean the real thing: architecture design, pretraining runs, custom CUDA kernels, loss functions from first principles.

My primary focus is on three areas: large language models (transformer architectures, training dynamics, RLHF, efficient inference), diffusion models (score-based generative modeling, latent diffusion, flow matching), and multimodal systems (vision-language alignment, cross-attention, contrastive learning). I care about understanding why things work, not just that they work.

What you'll find here

This blog is a research notebook made public. I write about:

  • Model architecture — dissecting transformers, diffusion U-Nets, SSMs, MoEs, and whatever comes next. Not summaries of papers; actual derivations and reimplementations.
  • Training at scale — mixed precision, gradient checkpointing, distributed training strategies, instability debugging. The parts of ML that live between the paper and the working model.
  • Mathematical foundations — variational inference, score matching, information theory, optimization geometry. I think math belongs in engineering writing.
  • Systems and tooling — PyTorch internals, custom autograd, CUDA programming, inference optimization. The machinery underneath the research.

I write when I notice something that took me longer to understand than it should have, or when I've done something that I wish I could have found written down somewhere.

About this site

Built from scratch with Next.js 15 (statically exported), TypeScript, Tailwind CSS, and JetBrains Mono everywhere. Markdown is processed with unified / remark / rehype; syntax highlighting comes from Shiki; math is rendered with KaTeX. There is no CMS. Posts are markdown files in a folder. The build reads them, renders HTML, and ships static files to GitHub Pages.

Source on GitHub.

Support

If my writing has helped you understand something faster, you can support more of it:

Thanks for reading.