My Projects

Google ADK Multi-Agent Framework with RAG, SOP, and BigQuery Agents

A multi-agent system on Vertex AI using Google’s ADK, featuring a manager that delegates to specialized agents: a RAG agent for document retrieval, an SOP agent for procedures in Google Sheets, and a BigQuery agent for natural language to SQL queries. It showcases how multi-agent orchestration can unify knowledge retrieval, procedural guidance, and data analytics.

Google Cloud ADK Multi-agent Framework Vertex AI RAG BigQuery

Automated Workflow for Electronics Data Extraction

An automated pipeline for extracting structured data from unstructured electronics datasheets (PDFs). The system combines OCR, markdown conversion, chunking, and LLM-driven parsing in a graph-based workflow to capture manufacturer part numbers, descriptions, packaging, and markings. It includes intelligent filtering, retries, and validation, producing clean, deduplicated outputs for large-scale electronics data processing.

Langchain Sequential Workflow OCR Open-source Tools Data Extraction

RAG Chatbot for Government Documents

A Django-based Retrieval-Augmented Generation (RAG) chatbot designed for government ordinances and resolutions. It integrates AWS Textract for document parsing, Weaviate for vector search, and the OpenAI API for generation, enabling natural language queries with accurate, source-grounded answers to improve access to government information.

RAG AWS Textract Weaviate Vector DB Open AI API Django Framework

My Resume Chatbot

This is an interactive chatbot that answers questions about my career, skills, and experience. Built with Gradio and powered by OpenAI’s LLM, it uses my resume to provide accurate responses. The chatbot also records user details (name, email, notes) and logs unanswered questions. It can be run locally with a simple setup or accessed directly via Hugging Face Spaces.

Interactive Chatbot Open AI LLM Gradio Hugging Face AI Agent

Multimodal RAG Pipeline for Document Analysis

This is an end-to-end system for processing PDF documents by extracting text and figures, generating AI-based figure summaries, and enabling hybrid retrieval across text and images. The pipeline stores figure summaries in a vector database while keeping the original images locally, allowing queries to return both textual and visual context for more comprehensive answers.

Multimodel RAG AWS Textract Chroma DB Open AI API Hybrid Search

AI-Driven Refinery Hydrogen Consumption Forecasting

This project enhances hydrogen consumption prediction in refinery hydrotreating processes using a machine learning pipeline with outlier detection, feature selection, and advanced regression modeling. Trained on refinery time-series data, the solution provides a more accurate and robust tool for optimizing hydrogen usage and improving refinery efficiency.

Auto ML EDA Dim Reduction Hyperparameter Tuning Ensemble Learning

Grocery Item Detection and Segmentation System

Designed and deployed an image detection and segmentation system for grocery items using YOLO and integrated with Streamlit for user interface.

YOLO Streamlit Computer Vision Data Agumentation Data Annotation

Classical ML for Filipino Fake News Detection

This project tackles fake news detection in the Filipino language by leveraging classical machine learning techniques instead of complex transformer-based models. Using TF-IDF for feature extraction and scikit-learn classifiers, the study evaluated individual, ensemble, and neural network approaches. The best-performing model was deployed through a Flask web app, providing a simple, efficient, and accessible tool to combat misinformation in the Philippines.

Classification ML Bag-of-Words TF-IDF Scikit-learn Hyperparam Tuning

Adaptive In-Context Learning for Niche Translations

This project explores the translation capabilities of CuatroLLM, a 1.3B parameter, four-language model (English, French, Spanish, German), with a focus on adaptive fine-tuning for specialized domains and low-resource language extension. We replicated CuatroLLM’s benchmark performance, improved English–French scientific translations using adaptive machine translation techniques, and tested English–Filipino translation. Results show that adaptive in-context learning boosts domain-specific translations in supported languages, but extending to low-resource languages like Filipino requires larger datasets and pre-training support.

PEFT/LoRA Hugging Face Benchmark Replication Adaptive MT

Tube Temperature Monitoring App

Reformer Tube Temperature Monitoring Application is a Dash-based web tool that connects to a PostgreSQL database for monitoring and managing reformer tube metal temperatures and air register openings. It provides data entry, visualization, reporting, user management, and revision tracking to streamline plant operations.

Python Dash PostgreSQL Database Management

Llama 3.2 Benchmark Evaluation

This project replicates Meta’s published results on the Infinite Bench benchmark, which tests the long-context comprehension of large language models. Using the Llama 3.2 1B Instruct model, the project runs evaluations on the longbook_qa_eng dataset and reports ROUGE-1 scores to measure performance.

LLM Benchmarking Llama 3.2 1B Hugging Face