Enterprise resource planning platform

A single system that brings a company's core operations together, from people and shifts to purchasing and vendors.

• Organization and HR modules, including employee management.
• Holiday and shift roster management.
• Procurement workflow with PRFs, vendors and item management.
• Automated build and deployment with CI/CD.

Built with Next.js, Spring Boot, Java, CI/CD

AI-powered pharmacy management system

Pharmacists scan a medicine's image and the system adds it to inventory or the point of sale automatically, so there is no manual data entry.

• Real-time image recognition pipeline that classifies medicine products.
• Inventory and sales workflow where stock updates automatically after each AI classification.
• Interface designed for fast-paced pharmacy counters.

Built with Image classification, React, Next.js, REST APIs, Supabase

Enterprise platform

A multi-role system for HR, managers and employees, each with their own access controls and workflows.

• Photo-based facial attendance that replaces manual clock-ins.
• Leave management, task and project tracking, and contract-based employee records.
• Real-time messaging and broadcasts for internal communication.
• Salary audit tools for payroll review, exception tracking and compliance reports.

Built with React, Next.js, Prisma, PostgreSQL, Neon, WebSockets

Deep learning, image classification

Classifies photos of waste into 12 categories, including plastic, metal, paper, three kinds of glass, clothes and shoes. It compares CNNs built from scratch against transfer learning.

• Trained on about 15,500 images from a public Kaggle dataset, split 80/10/10 into train, validation and test sets, with class weights to handle the imbalance between categories.
• Built a baseline CNN and a deeper CNN with batch normalization and dropout, then compared the Adam and SGD optimizers and ran an ablation study on dropout.
• Fine-tuned MobileNetV2, pre-trained on ImageNet, and reached 94.7% test accuracy, up from 72.4% for the best CNN trained from scratch.

Built with Python, TensorFlow/Keras, CNN, MobileNetV2, Transfer learning, scikit-learn

Natural language processing

Reads a tweet and labels it as hate speech, offensive language or neither, on a dataset where hate speech is only about 6% of the examples.

• Cleaned and tokenized about 24,800 tweets, and handled the heavy class imbalance with class weights and a stratified split.
• Compared a TF-IDF and logistic regression baseline against a simple RNN, a BiLSTM and a BiLSTM using pre-trained GloVe Twitter embeddings.
• Judged models on macro F1 instead of accuracy, since always guessing the majority class already scores 77%. The baseline led on macro F1 (0.74), while the GloVe model caught the most hate speech (about 73% recall).
• Analyzed misclassified tweets and built an interactive notebook where you type a tweet and see each model's prediction.

Built with Python, TensorFlow/Keras, scikit-learn, TF-IDF, BiLSTM, GloVe

Learning projects

Image classifiers and text models built to understand the fundamentals rather than lean on high-level wrappers.

• CNN image classifiers written from scratch.
• Transfer learning to improve accuracy and cut training time on domain-specific data.
• Text classification with tokenization, feature extraction and model evaluation.
• Distributed data processing with PySpark and Hadoop.

Built with Python, TensorFlow/Keras, NLP, PySpark, Hadoop

Two-actor pickup platform

Residents request a waste pickup and collectors get an SMS right away, with no mobile app needed.

• Full request lifecycle from submission to real-time SMS dispatch.
• ClickSend SMS API integration for reliable collector alerts.

Built with HTML, CSS, PHP, MySQL, ClickSend API

Web development

Responsive, fast, accessible websites delivered for real clients.

• otec.com.np – Corporate website for Otech Pvt Ltd
• pci.com.np – Corporate website for P.City Tech Nepal
• barbyjohn.com – Business landing page