Yubraj Sigdel — Academic Portfolio

Currently

What I'm working on now

As of 2026, I'm building a multilingual toxicity moderation system that classifies content across six categories of harm, with an emphasis on consistent behavior across languages and on decisions that trace back to a specific policy rather than an opaque score. A recurring question is how well these classifiers hold up on lower-resource languages and adversarial or out-of-distribution content.

Research interests

What I work on, and why

My research centers on making AI systems trustworthy and usable in settings where trust, language, or data are scarce. I'm particularly drawn to AI content moderation — building systems that flag misaligned content reliably while remaining auditable — and to explainable AI more broadly, since a moderation, medical, or safety-critical decision is only as useful as the reason behind it.

A second thread is NLP and multimodal LLMs for low-resource languages, particularly Nepali: OCR and handwritten-script digitization, semantic search, and Romanized-to-Devanagari transliteration are all instances of a broader question — how do we bring the same quality of language technology to languages with a fraction of the training data and tooling that high-resource languages have?

I'm also increasingly interested in AI for medical/healthcare applications, where the explainability and low-resource-data constraints I already work with intersect directly with real consequences for patients and clinicians. I'm open to research collaborations and conversations across any of these areas.

Education

Degrees & coursework

Bachelor's in Computer Engineering Kathmandu Engineering College, 2018–2023.
82.41%

Coursework Artificial Intelligence (AI), Mathematics (I, II, III), Applied Mathematics, Probability & Statistics, Numerical Methods, Discrete Structure, Big Data Technologies, Data Mining, Database Management System (DBMS), Digital Signal Analysis & Processing, Operating Systems (OS), Data Structures & Algorithms (DSA), Theory of Computation (TOC), Object-Oriented Programming (OOP)
High School Gorkha United Public School, 2016–2018.

Also

Research skills & methods

Research areas AI Content Moderation, Explainable AI (XAI), NLP, Multimodal LLMs, Low-Resource Language Modeling, AI for Medical/Healthcare, OCR & Handwritten-Script Digitization, Semantic Search, Transliteration
Methods & models ViT, NepBERT, mT5, YOLOv8, RoBERTa, Sentence Transformers, Fine-Tuning
Tooling PyTorch, Hugging Face, Pandas, NumPy

Publications

Peer-reviewed & preprints

Yubraj Sigdel, Sujan Shrestha, Nabin Neupane. College Chatbot Using RASA. KEC Journal of Science and Engineering, vol. 8, no. 1, 2024.
The rapid advancement of AI and machine learning has driven decades of change across technology. One such transformative technology is the chatbot — a conversational AI that answers user queries, often to streamline and automate tasks in education. This paper presents a chatbot that responds to questions about a college and related information, served through an API built with the Streamlit framework. It is built on the open-source RASA framework, which combines two components: Rasa NLU, which infers intent and extracts entities from user input, and Rasa Core, which generates responses from a probabilistic model. The model is evaluated with a confusion matrix and a prediction-confidence distribution, and on testing reaches a best score of 1.0 for accuracy, precision, recall, and F1.

Kathmandu Engineering College — Dept. of Computer Engineering (Sigdel; Neupane, Assoc. Professor); Dept. of Electronics, Communication & Information Engineering (Shrestha, Assoc. Professor).
DOI Code

Low-resource NLP

Nepali language research & models

A curated line of work on bringing modern language technology to Nepali — OCR and handwritten-script digitization, sentence embeddings and retrieval, and transliteration — with models and datasets released openly.

Handwritten Devanagari Document Conversion ViT + NepBERT, YOLOv8, 2024.
A two-stage architecture for digitizing handwritten Devanagari: YOLOv8 for word-level detection, then a Vision Transformer encoder paired with a NepBERT/RoBERTa language model for recognition and correction, reaching a 9.08% Character Error Rate.
- TrOCR
- ViT
- NepBERT
- YOLOv8
- OCR
Model Code
Romanized Nepali → Devanagari Transliteration Fine-tuned mT5, 2024.
An end-to-end transliteration pipeline from Romanized Nepali to Devanagari, with the fine-tuned mT5 model and its training dataset released on Hugging Face.
- mT5
- Transliteration
- Hugging Face
Model
Nepali Sentence Transformer FuseMachines Nepal, 2023.
A sentence-embedding model for Nepali, trained on translated STS datasets over NepaliBERT baseline embeddings to support semantic similarity and retrieval in a language with little labeled data.
- Sentence Transformers
- NepaliBERT
- Semantic search
Model
Context-based Nepali News Recommendation FuseMachines Nepal, 2023.
A Nepali news recommender built from scratch — scraping, embedding pipelines, and deployment — using contextual article embeddings to surface related coverage.
- Recommendation
- Embeddings
- Nepali NLP
AI for Women Farmers — Nepali OCR & NER FruitPunch AI, 2024.
Annotated Nepali documents for OCR, NLP, and NER systems and collaborated with OCR teams to strengthen Nepali text identification for an applied agricultural-support project.
- OCR
- NER
- Annotation

Thesis & undergraduate projects

Major & minor projects

Network Intrusion Detection System Kathmandu Engineering College, 2023. Major project / thesis.
The rapid advances in the internet and communication fields have resulted in a huge increase in the network size and the corresponding data. With the advancement of internet over the years, the number of attacks over the internet have been increasing. An intrusion detection system (IDS) is one such tool that prevents the network from possible intrusions by inspecting the network traffic, to ensure its confidentiality, integrity, and availability. Machine Learning algorithms are applied in IDS in order to identify and classify security threats. This report explores the comparative study of various machine learning algorithms such as K-nearest neighbor, random forest, Adaboost based decision tree, and support vector machine for various applications such as to provide security to organizations as well as help organizations identify their bugs and error-prone areas in the network. The above-mentioned algorithms are used to detect and classify the attack type. The NSL-KDD dataset, taken from Kaggle, is used to train and test the system model.

Keywords: intrusion, confidentiality, Adaboost, NSL-KDD.
Code
Face Detection & Recognition for Criminal Identification Kathmandu Engineering College, 2022. Minor project.
The face is one of the easiest ways to tell one person from another, and this project turns face recognition into an investigation tool for the crime department. A face given to the system — by video capture or an uploaded image — is reduced to its facial features and compared against a database that holds each criminal's facial information alongside a brief profile: name, age, offense committed, fingerprint, blood group, and so on. A match displays the individual's full profile; no match shows “Face Unrecognized” and offers to add the image to the database. Recognition is built on the LBPH algorithm — one of the earliest, simplest, and most efficient — across four sections: a user interface for searching profiles and inserting images, a dataset section that inserts and updates records, a training section that trains on those records, and a detection section that identifies faces using the dataset and trained model. Pairing facial with other biometric features makes criminal profiling more systematic and reliable, since such features are far less likely to be mismatched.

Keywords: face detection, face recognition, LBPH, biometrics, criminal identification.
Code