Kantinan
Rujiraruangporn

An open-source Python library for modular financial data fetching and feature engineering, designed for quantitative trading workflows. • Designed a Facade + Strategy pattern architecture providing a unified API to switch between providers (Yahoo Finance, Tiingo, Alpha Vantage, FMP) with a single parameter change. • Built a BaseDataFetcher abstract layer handling input validation, DatetimeIndex normalization, and robust HTTP retries with exponential backoff via tenacity. • Supports multi-asset data types: equities, crypto, forex, news, and fundamentals (income statements, balance sheets, cash flows).

Built an end-to-end automated quantitative trading pipeline that generates multi-horizon predictions and sends daily signals via Email. • Developed a PyTorch Multi-Head LSTM architecture to simultaneously predict multiple holding periods using Walk-Forward Cross Validation to prevent future data leakage. • Automated daily execution and fast-training for live inference, automatically retraining on the latest data to capture recent market regimes. • Created a realistic backtesting engine factoring in transaction costs, slippage, and dynamic ATR Stop-Loss, powered by data from Alpha Vantage, Tiingo, and Google BigQuery (GDELT).

Developed an object detection and segmentation pipeline for medical images to identify polyps.

Developed a production-ready conversational AI assistant acting as a professional Perfume Sommelier using FastAPI and LangGraph. • Engineered a True Hybrid Search RAG pipeline combining Sparse (BM25) and Dense (Qdrant Vector) retrieval, merged via EnsembleRetriever and re-ranked with a BAAI Cross-Encoder. • Implemented an Agentic Firewall to intercept prompt injections, alongside strict Qdrant JSON payload filtering for negative constraints. • Orchestrated a ReAct Agent architecture with conversational memory and dynamic tool calling (DuckDuckGo) across a dataset of 660 fragrances, featuring Hot-swappable LLMs (Gemini, Llama-3).

Spearheaded comprehensive research evaluating 7 ML/DL architectures for Quantitative Trading across 7,350 Walk-Forward Analysis test cases. • Engineered robust features using GDELT Sentiment Analysis and macroeconomic indicators (VIX, SPY). • Implemented Dynamic Labeling (Percentile Dynamic, Triple Barrier Method) to prevent regime overfitting. • Achieved a maximum CAGR of 104.84% (LLY) and +73.72% (XOM) using Sequential Deep Learning (GRU/LSTM), significantly outperforming the baseline while strictly managing downside risk (Max Drawdown < 5%).

Architected a Retrieval-Augmented Generation (RAG) pipeline utilizing Llama 3.2 3B Instruct to deliver accurate, domain-specific medical advice.

Built NLP pipelines covering tokenization, stemming, lemmatization, word embeddings, and ML-based text classification. Applied Transformer/BERT models and explored LLM integration including Llama 3.

Achieved an R-squared of 0.81 by optimizing and deploying an XGBoost Regressor. Engineered a comprehensive data pipeline processing 930k+ TMDB records.

JavaFX desktop application for group travel planning. Features multi-role auth (Admin/User), event creation & joining, team formation, shared timetables, commenting system, and user profiles. Built using MVC pattern with CSV-based file storage. Contributed Login system, User model, Create/Edit Trip, Navbar component, Admin panel, and login tracking.