ML STUDENT · RESEARCHER

Aditya Singhal — building at the intersection of ML research and engineering.

I'm a student at Gati Shakti Vishwavidyalaya working on language models, agentic systems, and mechanistic interpretability. I build to understand.

View work Get in touch
About
Aditya Singhal

Who I am

I'm an AI & Data Science student at Gati Shakti Vishwavidyalaya, Vadodara, specializing in Transportation & Logistics. I build things from scratch — not to collect credentials, but to actually understand how they work.

My current obsession is efficient language models: making small models punch above their weight through architectural choices, not compute. I'm also interested in agentic systems and what happens inside transformer networks when they do what they do.

B.Tech — AI & Data Science Gati Shakti Vishwavidyalaya · 2025 – 2028
Projects

Things I've built

Small Language Model

Complete

A 100–300M parameter language model trained from scratch in PyTorch on consumer hardware (M4 MacBook Air, 16GB RAM). Incorporates architectural efficiency techniques from DeepSeek — MLA, DeepSeekMoE, RoPE, GaLore. Fine-tuned via SFT on a human-refined dataset and deployed as a Discord bot with RAG and agentic capabilities.

PyTorch Python RoPE MLA RAG Discord.py

Discord Moderation Bot

Complete

An agentic Discord moderation system built on a ReAct-style loop with a tool registry, structured JSON tool calling via Gemma, observation injection, retry logic, and async Discord architecture. Built deliberately from scratch to understand every layer of how agents actually work.

Python ReAct Gemma Discord.py Async
Research

Writing & papers

Mechanistic Interpretability & Ablations on a Self-Trained SLM

Preprint · 2024

Forthcoming

A single-authored study on mechanistic interpretability of a self-trained small language model. Includes ablations on architectural components to understand what each contributes to model behavior.

Experience

Where I've worked

May - June 2026

Intern

DRM Office, Indian Railways

Exposure to operational systems and logistics workflows at a divisional railway management office

Skills

What I work with

Technologies

Python PyTorch HuggingFace LangGraph Discord.py Git

Areas

Language Models Mechanistic Interpretability Agentic Systems NLP Retrieval-Augmented Generation
Contact

Get in touch

I'm always open to interesting conversations, collaborations, or just a good technical discussion.