Featured projects.

AskMe

Intelligent document Q&A platform

Leverages RAG architecture for complex multi-document analysis. Integrates ALIParse for advanced document structure parsing, semantic vector retrieval, and LLM synthesis to provide accurate, context-aware answers across DOCX, PDF, and HWPX.

VLMFusion OCR

Multi-engine OCR with VLM post-correction

Combines EasyOCR, PaddleOCR, and Tesseract with a Vision-Language Model post-correction layer for superior accuracy on degraded, complex, or low-resolution document images. Confidence-based engine selection + VLM error correction.

FormGen

AI form field extraction & auto-fill

Detects form fields in HTML / DOCX / HWPX via LLM-based document analysis, with RAG-powered "fill from documents" that retrieves values from uploaded references via semantic search & re-ranking. Parallel LLM processing with order-preserving concurrency.

SlideGen

AI presentation generator

Creates structured slide decks from a text prompt or PDF context. Supports multi-provider LLMs — OpenRouter, Together AI, Cohere, Google Gemini, HuggingFace, Azure OpenAI — with dynamic Unsplash backgrounds and adjustable slide count (3–30).

Medical Document Disease Extractor

RAG-based clinical NLP pipeline

Extracts disease information from clinical documents and research literature. Supports Single Pass, Double Pass, and RAG analysis modes. Qdrant vector store with HNSW indexing, Qwen VLM for visual understanding, WebSocket real-time progress.

Hwpx2html

HWPX → HTML conversion engine

Converts HWPX (Hangul Word Processor) documents into clean, Pandoc-friendly HTML, plain text, or structured metadata via a pure-JavaScript pipeline with no native dependencies. Faithfully reproduces tables, embedded images, footnotes, lists, drawing shapes, and inline formatting in the final output.

• parse → resolve-styles → render pipeline preserves full document fidelity
• Usable as a CLI, standalone script, or JavaScript API
• No native binaries — runs anywhere Node.js does

Html2hwpx

HTML → HWPX conversion engine

Lightweight pure-JavaScript engine with no native binaries. Supports headings, paragraphs, bold/italic/underline, tables with colspan/rowspan, nested lists, hyperlinks, auto-sized images, inline styles, and code blocks. Standalone script, CLI, or Node.js library.

Chart2Excel + Chart QA

Chart data extraction & question-answering

Converts chart images into structured Excel data and supports natural language queries over chart content. Secured 2nd Place at AI Grand Challenge 2023 (Document QA track).

Semantic Segmentation on Cityscapes

SegNet, U-Net & FCN — a comparative study

Extensive experiments with pre-existing models for semantic segmentation on the Cityscapes dataset. Investigated SegNet, U-Net, and Fully Convolutional Networks (FCN) with VGG-16 as the encoder, applying enhancement techniques and providing a comprehensive performance comparison.

Image Classification on CIFAR-10

A study of Deep CNNs

Comparative study of Deep Convolutional Neural Networks on CIFAR-10. Developed a custom CNN with an optimal layer configuration via Deep Cognition, applied traditional and modern techniques across multiple DCNN variants, and benchmarked their performance.

Intrusion Detection System

Comparative ML on NSL-KDD dataset

Applied various classifiers to network intrusion detection data to enhance detection capabilities. Conducted a comparative analysis across classifiers using the NSL-KDD dataset to evaluate and improve identification of network intrusions.