AI & ML Student Maharashtra, India

Shubham Pardule

I build tools for training language models and exploring legal documents.

Selected work

Get in touch

Open to work · building TrainAI

Selected work

From data to working models.

Featured / ML tooling

TrainAI

Small language models.
Your hardware. Every step visible.

Trains a small language model from scratch on hardware you already own — no training code to write. It measures your machine, plans what's actually trainable on it, then handles the tokenizer, data pipeline, checkpointing, and export end to end.

Python
PyTorch
CLI
Hardware planner

Explore TrainAI Read the walkthrough

The training workflow

01Prepare the corpus
Inspect, tokenize, and split the data.
02Measure & plan
Test what the hardware can actually train.
03Train & evaluate
Checkpoint, resume, and score held-out text.
04Export the model
Take the trained weights into the next workflow.

One command to start

trainai quickstart ./my-corpus --out my-first-model

Actively developing

Applied ML / Retrieval

Legal AI Analytics

A local pipeline for exploring legal documents: MiniLM embeddings and FAISS for similar-case retrieval, with explanations and evidence to inspect alongside predictions.

The retrieval pipeline

01
Legal documents
The source material
02
MiniLM embeddings
Represent documents for similarity search
03
FAISS retrieval
Surface relevant cases and evidence

Python · FastAPI · MiniLM · FAISS

Explore the repository

Deployed web project

TransmitFlow

Move files between devices from a browser. Pair with a code or QR, then transfer over WebRTC without storing the files on a server.

Connection recovery, chunked transfers, and clear progress states.

Next.js · TypeScript · WebRTC · Socket.IO

Case study Open app Source

About

Learning AI by building it.

Building TrainAI is how I’m putting my AI & ML studies into practice. It gives me concrete decisions to work through, from preparing data to evaluating a trained model.

One decision that shapes my approach: measure real training steps before recommending a model configuration. A plan should reflect what the machine can actually handle.

TransmitFlow is my strongest deployed project so far. Building its interface and transfer system gives me the full-stack experience to take my AI projects beyond experiments.

A short introduction

shubham@web: ~

$ whoami
shubham pardule — AI & ML student
maharashtra, india

$ ls projects/
trainai · local model training
legal-ai-analytics · document retrieval
transmitflow · browser file transfer