✦ AI/ML Engineer · ML Research Intern · Full Stack Developer ✦

Naman
Parmar

// Training LLMs · Building RL agents · Shipping full-stack products

CS undergrad from Charusat · Built Gujarati GPT-2 & LLaMA2 from scratch · 295+ real users on PromptCraft · seeking ML & SWE internships

scroll
NP
Profile picture
295+ users
organically 🎉
Gujarati GPT-2
from scratch!

Hey, I'm Naman

I'm a CS undergraduate from Charusat University, Anand, India - building intelligent systems at the intersection of NLP research and full-stack engineering.

My core work is in low-resource NLP - I've designed and trained GPT-2 and LLaMA2 architectures from scratch for Gujarati, built custom SentencePiece tokenizers, and fine-tuned models on IndicSQuAD for generative QA. I also build RL environments and train agents using PPO.

On the engineering side, I ship full-stack products that people use - PromptCraft hit 295+ organic users in 14 months. I've interned as a full-stack developer and am comfortable across the Python/JS/React stack.

↳ always curious, always building

Projects

01
⚡ LLM from Scratch
Gujarati GPT-2

Built a 52M-param GPT-2 causal language model from scratch for the Gujarati language. Trained a custom SentencePiece tokenizer with 64K vocab on a 3.4GB corpus.

↳ Val loss: 5.84 · FP16 + AdamW · 5 epochs
PyTorch HuggingFace SentencePiece FP16
View on HuggingFace →
02
🔬 Fine-tuning
Gujarati QA Model

Fine-tuned the Gujarati GPT-2 base on IndicSQuAD for generative question answering. Used a structured Context / Question / Answer format prompt template.

↳ Val loss 5.84 → 4.80 (18% improvement) in 3 epochs
HuggingFace PyTorch IndicSQuAD
View on HuggingFace →
03
🏗 Architecture Design
Gujarati LLaMA2 Base

Designed a compact LLaMA2-architecture LM (6 layers, 512 hidden dim, 64K vocab) trained on 3.4GB of Gujarati text. Implemented custom LR scheduling with gradient norm tracking.

↳ 3.4GB corpus · Custom LR scheduler · Gradient norm tracking
PyTorch HuggingFace LLaMA2
View on HuggingFace →
04
🎮 Reinforcement Learning
Mars-Mars Game RL

Built a 2D platformer (Pygame) wrapped as a Gymnasium-compatible RL environment. Trained a PPO agent with shaped rewards handling dynamic wind, jetpack physics, and fuel management.

↳ PPO agent · Stable-Baselines3 · Custom shaped rewards
Python Pygame Gymnasium Stable-Baselines3
View on GitHub →
05
🚀 Live Product
PromptCraft

Deployed web tool powered by Google Gemini API generating tailored coding prompts. Built with JS + TailwindCSS - acquired real users organically with zero paid marketing.

↳ 295+ real users · 14 months organic growth
Gemini API JavaScript TailwindCSS
View Live →

Skills

AI & Machine Learning
PyTorch HuggingFace Scikit-learn LLM Fine-tuning Reinforcement Learning SentencePiece
Languages
Python JavaScript HTML / CSS C / C++
Frameworks & Libraries
Flask React Native TailwindCSS AngularJS
Tools & Databases
Git / GitHub Figma VS Code Postman MySQL MongoDB
Naman PyTorch HuggingFace RL / PPO Gujarati NLP Vision & YOLO - neural skill graph -

Experience & Education

JULY 2023 – AUG 2023
Full Stack Developer Intern
InnoBrain Technologies, Anand
  • Built an AngularJS frontend for a food store management system
  • Used TailwindCSS for responsive design; JSON Server for data interchange
2024 – EXP. MAY 2027
B.Tech. Computer Engineering
Charusat University, Anand
  • Currently in 6th semester · Focus on AI/ML systems & low-resource NLP
2021 – 2024
Diploma in Computer Engineering
B & B Institute of Technology, Anand
  • Built strong foundation in programming, systems & web development

Achievements

ML
🧠
Gujarati LLM Suite
Trained GPT-2 and LLaMA2 models from scratch for Gujarati - among very few open-source efforts for this low-resource language.
2023
🚀
295+ Organic Users
PromptCraft acquired 295+ real users in 14 months with zero paid advertising - through organic discovery and word of mouth.
2023
🔬
SSIP Hackathon
Participation in SSIP Hackathon Gujarat 2023, competing on innovative technology solutions for real-world problems.
INTERNSHIP
💼
InnoBrain Technologies
Full Stack Developer Intern - built production AngularJS frontend for a food store management platform with TailwindCSS.

Let's Build
Something Remarkable

Open to internships, research collaborations, and interesting problems. If you're working on something that matters - I'd love to hear about it.

parmarnaman19@gmail.com
↳ based in Anand, India · available for remote work
↑ top