AI Undergraduate | GenAI, Knowledge Systems & Software Engineering
Kathmandu, Nepal • LinkedIn • Email • GitHub

Undergraduate student pursuing a BSc (Hons) in Computing with Artificial Intelligence at Islington College (affiliated with London Metropolitan University).

Focused on building practical, grounded AI and software systems—specifically Large Language Model (LLM) applications, document understanding engines, and structured knowledge representation. Practical experience across Python backend architectures, local model orchestration, bilingual NLP/OCR pipelines, object-oriented Java systems, and deterministic web simulation engines.

• Languages: Python, Java, JavaScript (ES6+), TypeScript, HTML5/CSS3, SQL
• GenAI & Knowledge Systems: Local LLM Inference (Ollama), Open Knowledge Format (OKF v0.2), Dense Vector Retrieval (FAISS, sentence-transformers), Prompt Security (Injection Defense & Delimitation), Grounding & Contradiction Verification
• Document AI & NLP: Tesseract OCR (Devanagari & English), PyMuPDF (fitz), pdfplumber, Text Chunking & Normalization, scikit-learn
• Backend & Web Engineering: FastAPI, Next.js, React, Uvicorn, RESTful API Design, Node.js, Java Swing & AWT, Vanilla Web Standards
• Testing & Tooling: Pytest (TDD), Playwright E2E, Vite, Git, Linux/Bash, Microsoft Azure

Python 3.12, FastAPI, Ollama (qwen3:8b), FAISS, Tesseract OCR, Next.js 16, TypeScript, React 19

A local-first, offline-capable AI document-understanding platform engineered to help citizens navigate complex Nepali administrative forms, citizenship certificates, and government gazettes without cloud document leakage.

• Structured Knowledge Layer: Implements the Open Knowledge Format (OKF v0.2) to maintain a versioned, machine-validated concept graph for Nepali civil procedures rather than relying solely on naive chunk similarity.
• Trilingual Ingestion & OCR: Ingests PDFs and degraded scans using dual-engine Tesseract OCR (with dedicated Devanagari nep models) and PyMuPDF, dynamically selecting between Nepali and bilingual OCR modes.
• Multilingual Grounding: Employs paraphrase-multilingual-MiniLM-L12-v2 dense vectors in FAISS paired with a deterministic prompt-injection sanitizer and post-inference contradiction verification layer.
• Production Testing: Backed by 282 passing regression and pipeline tests and an interactive split-screen Next.js 16 workspace with live citation highlighting.

Vanilla JavaScript, Mulberry32 PRNG, WebAudio API, Vite, Playwright, Azure Static Web Apps

• Built a high-performance single-page sports simulation engine covering 13 major football competitions with zero runtime framework dependencies.
• Implemented a mathematical Poisson goal-distribution model paired with a seeded Mulberry32 PRNG for strictly deterministic, reproducible tournament simulations.
• Developed an interactive 2D pitch visualizer, custom 48-team World Cup bracket generator, procedural WebAudio sound synthesizer, and complete Playwright E2E test suite. Deployed live on Azure.

Java (JDK 17+), Java Swing, Object-Oriented Architecture, File I/O

• Designed and built a standalone desktop application simulating AI subscription plan tiers and prompt token quota management.
• Implemented object-oriented class hierarchies utilizing abstract base classes (AIModel), concrete polymorphic extensions (PersonalPlan, ProPlan), bounded team seat arrays, and transaction persistence.

Islington College / London Metropolitan University (Academic Research)

• Participated in comparative evaluation of lightweight open-weights LLMs (including Gemma 3 1B) across domain-specific Nepali datasets (Constitution & Law, civil procedures, socio-cultural texts).
• Investigated hallucination frequency and factual verification performance on multilingual benchmarks (HalluVerse and Poly-FEVER QA datasets).

• MedStore Pharmacy Inventory System: Modular Python CLI system with transactional integrity, inventory synchronization, and formatted text invoice generation.
• EcoMart Responsive Web Prototype: Multi-page responsive static e-commerce prototype with client-side cart state management and design wireframes.
• Spam Email Classifier: Introductory NLP experiment demonstrating text vectorization (CountVectorizer) and probabilistic Naive Bayes classification.

• Advancing evaluation methodologies for agentic workflows (tool use, multi-step reasoning, plan-verification loops).
• Scaling local inference latency optimization (quantization, vLLM / llama.cpp backends).
• Expanding Open Knowledge Format schemas for complex public-policy and regulatory documents.

• Email: omdangol68@gmail.com
• LinkedIn: linkedin.com/in/om-dangol-0a39243a7
• GitHub: github.com/omwe77