Ajaykumar S

NameGPT - Tamil Names Generator: Built a GPT-style Tamil name generator from scratch using PyTorch. Trained a 1M parameter model on 40,000+ Tamil names, achieving 95% accuracy. Implemented data preprocessing, tokenization, training, and inference pipelines. Developed an interactive inference application using FastAPI for Tamil name generation through a REST API.

Arduino

Applab AI - Software Development: Built a system where you can describe an Arduino project in plain English and get back working Python, C++ Sketch, and a web UI. It debugs itself until the app runs. Vibecoded the existing TypeScript Arduino codebase with Antigravity so the firmware can take instructions from the AI server and let agents call tools on the board.

Ran everything in Docker on AWS with Postgres for memory. Hit 94% success on 250+ test prompts and cut build time from hours to Transformers Tweak - Research: Modified the transformer's attention layer and added positional embeddings to the key and value vectors instead of just the input. Wanted to see if the model would track word order better.

Matched baseline loss/accuracy, but reduced attention entropy by 18% on 2k-token inputs and showed cleaner semantic clusters in attention maps.

AI Engineer - Outbell Communications Private Limited - Madurai

(2024-07 - 2026-01)

• NameGPT - Tamil Names Generator: Built a GPT-style Tamil name generator from scratch using PyTorch. Trained a 1M parameter model on 40,000+ Tamil names, achieving 95% accuracy. Implemented data preprocessing, tokenization, training, and inference pipelines. Developed an interactive inference application using FastAPI for Tamil name generation through a REST API.
• Arduino Applab AI - Software Development: Built a system where you can describe an Arduino project in plain English and get back working Python, C++ Sketch, and a web UI. It debugs itself until the app runs. Vibecoded the existing TypeScript Arduino codebase with Antigravity so the firmware can take instructions from the AI server and let agents call tools on the board. Ran everything in Docker on AWS with Postgres for memory. Hit 94% success on 250+ test prompts and cut build time from hours to
• Transformers Tweak - Research: Modified the transformer's attention layer and added positional embeddings to the key and value vectors instead of just the input. Wanted to see if the model would track word order better. Matched baseline loss/accuracy, but reduced attention entropy by 18% on 2k-token inputs and showed cleaner semantic clusters in attention maps.
• LawGPT - Finetuning: Scraped and curated 40k+ Indian law Q&A pairs from judgments, acts, and legal forums. Converted them into chain-of-thought format and built a FAISS vector store of 120k+ bare acts + case laws for grounding.

Finetuned

Gemma 2B with QLoRA + QAT, then added a LangChain RAG pipeline that retrieves relevant statutes before answering. Cut hallucinations on held-out questions by 42% vs finetune-only. 4-bit quantized model hits 74% answer accuracy and 97% of FP16 MMLU-Law score. End

B. Tech - Artificial Intelligence And Data Science - PSNA College of Engineering and Technology (2020-06 - 2024-03)