Language models, systems, research engineering
Aadit Suryawanshi
I am currently focused on building large language models from first principles. I spend much of my time experimenting with training pipelines, fine-tuning strategies, evaluation methods, and model architectures so I can understand their trade-offs directly.
Projects
LiteGPT
A clean decoder-only Transformer language model (~25M parameters) trained from scratch on a single NVIDIA A5000 GPU.
Recall-OS
RecallOS is an AI-native enterprise knowledge operating system that allows organizations to ingest, organize, search and reason over every piece of company knowledge.
SmolLM-135M_Med
End-to-end pipeline for continued pretraining, supervised fine-tuning, and evaluation of SmolLM-135M on medical datasets.
Xcal
Realtime Excalidraw
Current rabbit holes
- Training GPT-style language models from scratch in PyTorch
- Reinforcement learning environments
- Fine-tuning open-source LLMs for medical QA
Recent reads
About me
I like understanding systems beneath the abstractions. Rather than treating models and frameworks as black boxes, I enjoy implementing them from first principles to understand how they work and where their trade-offs lie.
My interests span the full AI stack—from training and evaluating language models to building distributed backend systems, retrieval pipelines, and full-stack AI applications. I'm particularly interested in turning research ideas into reliable, scalable products.
I learn by building. Most of my projects begin with a question I'm curious about and grow into end-to-end systems that combine machine learning, software engineering, and practical experimentation.
Featured article
Building a 25M parameter GPT from scratch
A practical write-up on architecture choices, data mixtures, tokenizers, training issues, and what I learned after training LiteGPT on FineWeb and TinyStories.
Read article