I understand the problem
before I write a single line.

I'm Jayaragul, an AI engineer and builder in Coimbatore, India. I work on LLMs, agentic systems and full-stack products, and I've learned that the hard part is almost never the code. It's knowing what you're actually solving.

See the work Let's connect
Jayaragul N, AI engineer based in Coimbatore, India
Jayaragul N Building
Programs & communities I've been part of
Google for Developers GDG Coimbatore IEEE NVIDIA AI Tamil Nadu The Exypnos

The instinct

Most people jump straight into building. I don't.

I start by asking a different set of questions: the ones that decide whether the thing you build will actually matter:

  • What is this problem really about?
  • What does the domain tell us?
  • What would a great solution actually look like?
  • What impact will the product deliver?

That instinct, to understand before acting, is the lens through which I approach everything I build.

The hard part was never the code.

Beyond engineering, I enjoy research, reading papers, and experimenting with new ideas, constantly exploring how systems work under the hood and trying concepts out in real implementations. That's how I ended up training a language model from scratch rather than just calling one through an API.

I also work across the full development lifecycle, from development to deployment. And beyond code, I've competed in hackathons, led ideation sessions, and organised tech events. Those sharpened not just my problem solving, but my ability to collaborate, communicate, and move fast with a team.

How I build

Technically, I'm grounded in Python, SQL, Git and full-stack development, with hands-on experience in Django, machine learning, OpenCV, data science, agentic AI and LLMs.

But I don't just apply tools. I design systems that combine backend logic, intelligent algorithms, and agentic flows that can plan, reason and execute tasks in structured steps — built for real production use cases, not demos.

Foundations

Python · SQL · Git · JavaScript · HTML & CSS

Backend & Full stack

Django · Flask · REST APIs · MySQL · MongoDB · Firebase

Intelligence

LLMs & fine-tuning · Agentic AI · Machine learning · Deep learning · NLP · OpenCV · Data science

Shipping

Google Cloud · Deployment · Linux RTOS · Motion profiling · Version control

0

Parameter language model, trained from scratch

0

Awards across hackathons, papers and ideathons

0

Developers in a community I helped grow from 12

0

Projects shipped across AI, vision and robotics

Selected work

A few things I've built. Each one started with a question about the domain, not a question about the framework.

0 shipped warming up
Operations research

A factory that tells you what's wrong with it

Factory simulators cost more than the machine you're deciding whether to buy. TwinOps is a free one that runs in a browser tab: describe a line in plain English, press play, and watch which station is actually limiting output.

The hard part wasn't the animation. It was measuring blocked and starved time as first-class state, because that's what lets it name your bottleneck instead of handing you utilisation numbers to interpret. Then it re-runs your line with each fix applied, so you get a decision rather than a diagnosis.

That panel is the actual discrete-event engine running in this page, seeded fresh each visit. Watch the queue fill in front of the slow station.

  • Discrete-event simulation
  • Python
  • Zero-dependency JS

Open the studio Source

27M parameters
Language models

A language model, built from scratch

Everyone can call an LLM API. I wanted to know what was actually happening inside one, so I built and trained a 27-million-parameter model from the ground up on the TinyStories dataset. Tokenizer, transformer blocks, training loop, text generation: all of it, by hand.

Working at that scale forces you to understand every decision, because there's nowhere for a mistake to hide. It changed how I read papers, and how I debug models I didn't write.

  • PyTorch
  • Transformer architecture
  • Tokenization
  • TinyStories
Drone in flight — subject of the micro-Doppler radar classification project
Signals & ML

Telling a drone from a bird

Radar sees a small moving object. Is it a drone or a bird? The answer lives in the micro-Doppler signature: the tiny frequency shifts created by spinning rotors versus beating wings.

I built a classifier that reads those signatures. The interesting work wasn't the model; it was understanding the physics well enough to know what the model should be looking at.

  • TensorFlow
  • Signal processing
  • Scikit-learn
Robotic arm built by Jayaragul, controlled from a servo control panel on a laptop
Robotics

Motion that doesn't jerk

A robotic arm that simply moves to a position will slam into it. Smooth motion needs motion profiling: planned acceleration and deceleration curves, running under real-time constraints on Linux RTOS.

Built the arm, the control panel, and the profiling underneath it. Hardware is honest feedback: it either moves properly or it doesn't.

  • Linux RTOS
  • Motion profiling
  • Python
  • IoT
Medical NLP project mapping described symptoms to likely conditions
Applied NLP

Reading symptoms like a clinician would

An NLP system that maps described symptoms to likely conditions and suggested treatment. The domain mattered far more than the architecture here. Medical language is full of overlap, negation and hedging that a naive model reads exactly backwards.

Alongside it, a pneumonia classifier for chest X-rays using OpenCV, and a full-stack system for patient records, appointments and prescriptions.

  • NLP
  • OpenCV
  • Python
  • Full stack

Attendance, automated

Camera-based attendance using face recognition with OpenCV, served through Flask.

Fine-tuning on private data

Adapting a compact language model to domain-specific data, so it works where a general API can't.

Open source

Contributions to machine learning and web development repositories.

Recognition

2026 · GDG Coimbatore

Winner — Tech for Good: Build with AI

Won Best Use of AI for Zero Hunger & Economic Growth at the Tech for Good hackathon hosted by GDG Coimbatore — part of Google Developer Groups, Google's global developer community program.

Beyond the code

I'm currently building in the AI space, scaling real products, and working toward becoming a recognised voice in the developer community. A lot of that happens away from the editor: teaching, organising, and getting people into the same room.

Writing

I write about what I learn

Notes on building AI systems, reading papers, and the things that only become obvious once you've shipped them. Published on Medium.

Read on Medium

Let's connect

If you're working on something meaningful in tech, I'd like to hear about it. I'm open to roles and collaborations in AI engineering, applied machine learning and full-stack product work.

jayaragul.in@gmail.com

Coimbatore, Tamil Nadu, India
Usually quickest to reach on LinkedIn.

Download résumé

Opens in your email app, addressed to me.