This dataset contains a comprehensive Question & Answer collection focused on water management technologies, irrigation systems, and related agricultural practices for sustainable farming. The dataset is derived from technical documentation and research publications related to water management in ag Read more
This dataset contains a comprehensive Question & Answer collection focused on plant diseases, their management, treatment protocols, and diagnostic techniques. The dataset provides detailed information about various plant pathologies, fungicide applications, and disease identification methods for ag Read more
This dataset contains a comprehensive Question & Answer collection focused on soil management, soil health, organic farming practices, and soil-related agricultural techniques. The dataset is derived from technical guides and documentation related to sustainable soil management practices, with empha Read more
Decoder-only transformer trained on TinyShakespeare, replicating the original OpenAI GPT architecture from scratch. Read more
Bidirectional encoder pre-trained with masked language modelling on the Cornell Movie Dialogs corpus. Read more
Cycle-consistent unpaired image translation on Cityscapes — two generators, two discriminators, cycle + identity losses. Read more
Differential attention replicated from scratch — two attention maps subtracted to cancel noise. Trained on TinyShakespeare on A100. Read more
LSTM-based Seq2Seq encoder-decoder for German→English translation. Train/val loss ~1.38 in 10 epochs. Read more
QLoRA fine-tuning scripts using PEFT + BitsAndBytes for both decoder and encoder-type models. Read more
From-scratch implementations of Bahdanau and Luong attention in PyTorch. Read more
Encoder-decoder transformer for English→Hindi translation on Samanantar (~25M params). Published on HuggingFace. Read more
Sparse MoE transformer replicated from scratch on TinyShakespeare. Train loss 2.04 / val loss 2.09 in 1,000 steps on T4. Read more
Direct Preference Optimization applied to Qwen0.5B-Instruct on UltraFeedback. Train loss 0.67 in 3,000 iterations. Read more
Reference-free preference optimization (SimplePO) on OPT-330M. Batch size 128, lr=2e-5, beta=2 on UltraFeedback. Read more
Low-rank adaptation implemented from scratch in PyTorch. Train/val loss ~3.5 in 1,000 steps on A100. Read more
Odds Ratio Preference Optimization on OPT-330M. Reference-free alignment reaching train loss 1.70 in 3,000 iterations. Read more
Google’s Gemma architecture replicated from scratch — multi-query attention and GeGLU activations on TinyShakespeare. Read more
Contrastive vision-language model trained on Flickr8K. Train loss 1.3 / val loss 2.2 in 30 epochs on T4. Read more
Llama trained with PyTorch DistributedDataParallel (torchrun). Val loss 1.1 in 8,000 iterations on TinyShakespeare. Read more
Visual instruction tuning replicated from scratch on Flickr8K. Train loss 0.23 / val loss 0.22 in 5 epochs on T4. Read more
GRU-based Seq2Seq with both Bahdanau and Luong attention from scratch. 128 hidden units, 50 epochs. Read more
Whisper ASR from scratch — CNN on 80-channel mel spectrograms + 6-layer transformer decoder. Trained on GigaSpeech. Read more
LSTM from scratch (~128K params). 128 hidden units, 50 epochs. Train loss 0.49 / val loss 0.48. Read more
90M-parameter Gemma 3 with local sliding-window attention (128-token blocks). Val loss 1.77 in 25k steps on TinyStories. Read more
1.2B-parameter MoE (32×12M experts, top-1 routing) trained on TinyStories. Val loss 1.70 in 20k steps on Kaggle P100. Read more
Compact transformer ASR (288-dim, 6 heads) trained on GigaSpeech for 1,500 steps. Notes on overfitting at ~25 hours. Read more
Conditional GAN for paired image-to-image translation (aerial→map) replicated from scratch. PatchGAN discriminator. Read more
Sigmoid-loss vision-language pretraining replicated from scratch on Flickr8K — avoids global softmax normalisation. Read more
Tacotron-style transformer TTS from scratch — 512-dim phoneme encoder, mel spectrogram decoder, 16kHz on GigaSpeech. Read more
VAE on CelebA (128×128). 4-layer conv encoder, 32D latent, ConvTranspose decoder. Reconstruction + KL loss over 200 epochs. Read more
Wasserstein GAN and WGAN-GP implemented from scratch on MNIST — gradient penalty for stable training. Read more
DeepSeekV3-inspired MoE with latent attention trained with Muon optimizer. Pre-trained weights on HuggingFace. Read more
Conditional GAN on MNIST — class-conditioned 64×64 digit generation. 30 epochs, BCE loss, TensorBoard logging. Read more
Contrastive Language-Audio Pretraining from scratch on GigaSpeech. 768D text / 2048D audio → 1024D shared space. Read more
Deep Convolutional GAN trained on CelebA and CIFAR-10. ~7,800 steps (CelebA) and ~11,700 steps (CIFAR-10). Read more
16×4 MoE with Multi-head Latent Attention and auxiliary-free load balancing, trained on TinyStories on Kaggle P100. Read more
Short description of portfolio item number 1 Read more
Short description of portfolio item number 2 Read more
| Movies Review System | Spoiler-Free Sentiment-Analysis based Movies Review System) |
Introducing the Movie Review System, where AI meets movie magic to revolutionize how viewers experience films. This project goal is to provide an interface for spoiler-free reviews and sentiment analysis, enhancing the viewing journey. With advanced models like Voting Classifier and Bi-LSTMs powered by Keras and TensorFlow, we achieve impressive metrics—a 91% accuracy, 91% precision,... Read more
| MoviesMania (Geek-o-thon) | A Reverse Search based Movies Recommendation System |
Step into the future of entertainment discovery with MoviesMania. The rpoduct aims to simplify your search for the perfect movie or web series. Using various AI/ML techniques and elements, we analyze uploaded video clips to predict movie titles and recommend similar content with an impressive accuracy. Experience flavoured recommendations tailored to your tastes, powered by... Read more
| PlogPayouts | AI-driven Plogging System |
Transform your daily jog into a mission for a cleaner world with PlogPayouts. Our innovative website + app rewards you for collecting litter, promoting fitness and environmental cleanliness. Utilizing AI for trash categorization and optimized routes, and fostering community through shared stories, PlogPayouts turns every step into a step towards a greener, more inclusive society.... Read more
| Insight-Ed (HackNITR 5.0) | EdTech Platform for Student and Teacher |
Imagine an online classroom where teachers instantly know when and why students lose focus. Our AI-powered solution bridges the knowledge gap by detecting student emotions and attentiveness, highlighting problem areas, makes the teacher aware of each student’s progress. With features like reverse video search, dynamic questionnaires, and advanced Q&A bots, we transform the learning experience,... Read more
| FarmGenie (GeoHack 2024) | Empowering farmers with real-time insights and expert guidance via AI-driven space |
Our platform utilizes LLMs and a Mixture of Expert (MoE) approaches to provide precise guidance on soil management, plant disease identification, and irrigation techniques. Built as a scalable web application with a Next.js frontend and backend, and supported by a Redis queue and multiple worker nodes, FarmGenie ensures robust performance. The system’s multilingual support, interactive... Read more
| NeatRL Playground | AI Games Showcase powered by Reinforcement Learning |
Beautiful, interactive website showcasing AI-powered games with reinforcement learning agents. Features Pong AI with Deep Q-Learning, real-time WebSocket communication, and smooth animations. Deployed on Vercel (frontend) and Render (game server) with production-ready health checks and headless mode. Read more
| Paper Replications | ML/DL Research Paper Implementations |
A comprehensive collection of code implementations replicating results from influential machine learning and deep learning research papers. Features 30+ models including Transformers, GANs, Vision models, and RLHF techniques. Read more
| NeatRL | Deep Reinforcement Learning Algorithms Library |
Comprehensive implementations of deep RL algorithms including DQN, A2C, PPO, DDPG, TD3, and SAC. Features one-file implementations, experiment tracking with W&B, automatic video recording, and support for Gymnasium environments. Main NeatRL library provides high-quality training utilities with focus on simplicity and performance. Read more
| SmolCluster | Distributed Deep Learning Library for Heterogeneous Hardware |
Educational Library for training/inference of neural networks across heterogenous compute like Mac minis, Raspberry Pi, and GPUs, written using only socket library in Python. Supports FSDP, Classic Data Parallelism, Elastic DP, Synchronous Parameter Server, and Model Parallelism. Read more
Tabular Q-Learning and Value Iteration implemented from scratch as educational notebooks. Read more
DQN agent trained on Flappy Bird using pixel observations, experience replay, and epsilon-greedy exploration. Read more
DQN agent trained on VizDoom Basic via Gymnasium wrapper, with grayscale preprocessing, replay buffer, and W&B logging. Read more
Group Relative Policy Optimization — DeepSeek-R1’s RL training objective implemented from scratch. Read more
Implementation of DQN-FrozenLake reinforcement learning algorithm Read more
Implementation of FlappyBird-PPO reinforcement learning algorithm Read more
Implementation of Frozen-Lake reinforcement learning algorithm Read more
Implementation of Imitation Learning reinforcement learning algorithm Read more
A PyTorch implementation of a Mixtral inspired transformer model with Mixture of Experts (MoE), designed for text generation and understanding tasks. This model is built on the Mixtral architecture wi… Read more
A compact implementation of an Encoder-Decoder Transformer for sequence-to-sequence translation tasks. This project implements a translation model from English to Hindi using the Samanantar dataset. Read more
A PyTorch implementation of a DeepSeek V3 inspired transformer model with Mixture of Experts (MoE), Latent Attention, and other advanced features. Read more
So, I trained a Llama a 88M architecture I coded from ground up to build a small instruct model, going through the below-mentioned stages from scratch. Read more
A PyTorch implementation of a Mixtral inspired transformer model with Mixture of Experts (MoE), Flash Attention, and other advanced features. Read more
So, I trained a Llama a 130M architecture I coded from ground up to build a small instruct model, going through the below-mentioned stages from scratch. Read more
Undergraduate course, University 1, Department, 2014
This is a description of a teaching experience. You can use markdown like any other post. Heading 1 Heading 2 Heading 3 Read more
Workshop, University 1, Department, 2015
This is a description of a teaching experience. You can use markdown like any other post. Heading 1 Heading 2 Heading 3 Read more
Hugging Face