Ajay Kumar

On-Device LLM Systems: From Model Conversion to Offline Mobile Inference (Qwen2 + GGUF + Flutter).

Built an end-to-end edge AI pipeline to run Large Language Models fully offline on mobile devices, from model conversion to real-time inference in a Flutter app. Developed a custom pipeline to convert Qwen2 models (.safetensors) into GGUF format for compatibility with llama.cpp, with full control over tensor mapping and tokenizer integration. Models were then quantized to optimize performance for resource-constrained devices. Integrated llama.cpp into a Flutter-based chat application to enable real-time, on-device inference with no API calls, no internet dependency, and complete privacy.

Akkadian to English Translation using ByT5 + LoRA (Kaggle competition)

Built a neural machine translation system for Akkadian (ancient, low-resource language) to English as part of the Deep Past Challenge (Kaggle). The project focuses on handling extreme data scarcity, noisy labels, and unknown scripts using modern parameter-efficient fine-tuning techniques.

Vision Transformer & MAE for Semantic Image Retrieval.

Built a semantic image retrieval system using a Vision Transformer encoder trained with Masked Autoencoder pretraining on the Food-101 dataset. The encoder generates high-dimensional image embeddings, which are used to retrieve visually similar food images through similarity search. Additionally, experimented with a Transformer-based caption generation pipeline using cross-attention, highlighting practical challenges in aligning visual features with natural language.

हिंदी GPT Model from Scratch — From Pretraining on 1.35B tokens to SFT & Deployment using PyTorch.

Designed and trained a decoder-only GPT-style language model from scratch in PyTorch, implementing masked multi-head self-attention with Rotary Positional Embeddings (RoPE), feed-forward networks, residual connections, normalization, and a stacked decoder architecture from first principles. The model was pretrained on ~1.35B Hindi tokens using a custom SentencePiece BPE tokenizer (Hindi-only vocabulary), enabling experimentation with large-scale language modeling, convergence stability, and efficient training on limited hardware.

Implementing the Transformer Encoder from scratch using PyTorch.

Implementing the Transformer Encoder from scratch using PyTorch — exactly as described in “Attention is All You Need” and also implements Multi-Head Attention, Positional Encoding, Feed-Forward networks, Residual+LayerNorm, and an encoder stack from first principles. The encoder was trained on KP20k (BIO labels) to validate the implementation and to experiment with attention behavior, head effects and extraction heuristics.

YOLOv1-Research Paper-Implementation-using-PyTorch-from-scratch.

This project is a from-scratch implementation of the YOLOv1 (You Only Look Once) object detection paper using PyTorch. I implemented the entire pipeline — architecture, loss function, dataset parsing, and model training — to deeply understand how YOLO works at its core.

Implementation of GAN Research Papers: From Concept to Face Generation.

Implemented and trained a DCGAN model to generate realistic human faces using the CelebA dataset. Built an interactive web interface with Flask, HTML, CSS, and JavaScript that allows users to generate faces, morph between two faces with a slider, and create batches of 100 synthetic faces. Deployed the project on Hugging Face Spaces, showcasing end-to-end GAN implementation from research paper to real-time user interaction.

AI-Powered Stock Market Insights: Building a Chrome Extension with Selenium, FastAPI, and LangChain.

I built a Chrome extension powered by Selenium, FastAPI, and LangChain that works as an AI-driven stock market chatbot assistant. With a simple natural language query, it can fetch real-time stock performance, show detailed stats for any stock, provide sector-wise insights, pull the latest market news, and display interactive charts — all directly inside the browser.

Implementation of U-Net Research Paper using PyTorch for Car Image Segmentation and Mask Generation.

I have implemented the original U-Net research paper using PyTorch and then trained this model on a car dataset. created the frontend to use this model and generate the mask of the car. For the frontend, I used Flask, HTML, and CSS. In the app, I can upload an image, and by clicking a button, it generates the mask and overlay of the original image.