Portrait of Rohit Bele

About

Hi, I’m Rohit. I’m a machine learning engineer interested in understanding how models learn, reason, and generalize. I’m working toward LLM research engineering by implementing models, checking the mechanics, and documenting what the experiments actually show.

Professional background

I hold a B.Tech in Computer Science and Engineering, with a specialization in Data Science and Machine Learning. My professional experience spans data engineering, applied AI systems, and software engineering, with hands-on work in Python, SQL, ML pipelines, and AI applications.

What I’ve built independently

My personal projects are a way to inspect the machinery directly:

  • Seq2Seq from Scratch: explicit LSTM gates and a small-scale English–French translation study. The experiments report low BLEU scores and no consistent benefit from source reversal.
  • Micrograd learning extensions: correctness fixes, gradient tests, and a backpropagation walkthrough built on Andrej Karpathy’s scalar autograd engine.
  • A character-level RNN: a NumPy language model with manual backpropagation through time, validation tracking, and sequence generation.

The project case studies connect implementation choices to results, limitations, and source code. These are independent projects, separate from my professional work.

What I’m studying now

My current focus is PyTorch, Transformers, LLM training and evaluation, and reinforcement learning from human feedback (RLHF). Training dynamics, reasoning, and generalization are longer-term research interests. I use this notebook to work through them; they are areas of study rather than claims of completed research.

Beyond the models

rohitwrites is also where I think out loud about engineering, learning, and life. Books, philosophy, and personal writing belong here alongside the technical notes.

Elsewhere