Vansh Visariya

Gen AI & Agentic AI Engineering Intern, EY

Last updated: June 2026

This work presents the practice of Vansh Visariya, an AI & ML engineer building scalable agentic systems. Currently interning at EY, the author’s work spans distributed federated learning, LLM-based automation, and transformer architectures built from scratch. The following sections detail methods, selected projects, and ongoing work.

Figure 0. Training convergence of the author’s practice over time — loss decreases monotonically with minor oscillations, finding its minimum in recent epochs.

§1. Introduction

The author is currently a Gen AI & Agentic AI Engineering Intern at EY (May 2026–Present), collaborating with senior AI consultants and architects to design, deploy, and optimise scalable AI systems. Prior to this, he worked as an AI Agent Intern at IPAGE UM Services (Oct 2025–Jan 2026), where he built intelligent automation agents for email replies and social media workflows using n8n, LLMs, and vector databases — achieving 2–3× faster content creation.

The author holds a GPA of 9.1/10 at SVKM’s Dwarkadas J. Sanghvi College of Engineering, Mumbai University, and is expected to graduate in 2027. His work has been recognised as a finalist at Quasar 4.0 (National Level Hackathon, Jan 2026), among the top 10 teams in Megahack 5.0 (out of 250 teams, Mar 2025), and a finalist at the Innovize Hackathon at VJTI, Mumbai (Feb 2025).

§2. Methods

All experiments are conducted in Python with PyTorch, TensorFlow, and Keras. Production deployments use FastAPI serving layers and Docker containerisation. Vector search workflows leverage ChromaDB and Pinecone. Agentic pipelines are orchestrated with LangChain, LangGraph, and n8n.

Languages

PythonC/C++SQL

Frameworks & Libraries

FastAPIPyTorchTensorFlow KerasOpenCVLangChain LangGraphn8n

AI/ML Tools

HuggingFaceMLFlowDocker Airflow

Databases & BI

SupabaseMySQLPinecone MongoDBPowerBI

Certifications

TensorFlow for Deep Learning Bootcamp — end-to-end theory, practice, and projects in ML, DL, and NLP.
Data Science, Machine Learning & Deep Learning — full certification with hands-on projects.

§3. Selected Work

Figure 1. ASTRA — A production-ready distributed Federated Learning platform with privacy-preserving model training, async aggregation, and a real-time full-stack dashboard.

PyTorchFastAPIDockerNext.jsFederated LearningDP-SGDLoRA

Repository ↗

Figure 2. Kaito-AI — An intelligent AI assistant using LangChain, LangGraph, and LLMs for real-time web search and RAG with semantic vector search.

LangChainLangGraphStreamlitChromaDBRAGLLMs

Repository ↗

Figure 3. KaitoModel — A GPT-2 inspired transformer built from scratch in PyTorch — masked multi-head self-attention, positional embeddings, causal masking, and a full training pipeline.

PyTorchTransformersGPT-2NLP

Repository ↗

§4. Writing

The author maintains a technical blog covering topics in production ML, agentic systems, and the craft of engineering AI.

Read the work →

New writing appears here and at /writing. The blog is typeset with the same rigour as the rest of this document — justified text, hyphenation, figure captions, and margin notes where warranted.

§5. Acknowledgements

The author would like to thank anyone who read this far. For collaboration, opportunities, or to argue about model architectures — reach out below.

Email: visariya.vansh@gmail.com

GitHub: github.com/vansh-visariya

LinkedIn: linkedin.com/in/vanshvisariya