Projects

Portfolio projects by Muhammad Iqbal Hilmy Izzulhaq across AI agents, RAG, MCP tool servers, NLP, LLM evaluation, full-stack AI apps, and applied data science.

VerdictBench: Benchmarking Long Context and Retrieval-Augmented Generation on Indonesian Constitutional Court Verdicts - An empirical benchmark comparing Long Context, Dense RAG, and Multi-Stage RAG on 50 Indonesian Constitutional Court verdicts and 300 human-reviewed QA pairs. Phase 2 shows Long Context and Dense RAG are statistically tied on gold-evidence faithfulness across Gemini 2.5 Flash and GPT-4o Mini, while...
EPC Tender Screening MCP - An MCP-only agent tool server for evidence-based EPC tender screening. It exposes HTTP and stdio MCP transports so Claude, ChatGPT, and other MCP-compatible agents can upload company profiles and tender text, extract requirements, match evidence, generate compliance matrices, and produce human-revi...
T-cell Evidence Workbench - An evidence-first bioinformatics workbench for auditing and prioritizing primary human CD4+ T-cell CRISPRi Perturb-seq results. A deterministic Python pipeline audits 33,983 perturbation-condition rows across four donors and three culture conditions, applies six registered source-QC gates, and prod...
Distill – AI Paper-to-Insights - A full-stack AI web app that transforms academic papers into 7 interactive output formats: infographic, quiz, flashcards, mind-map, data table, slides, and report. Upload a PDF or paste an arXiv link; Gemini 2.5 Flash reads the paper natively (no text extraction) and generates structured, renderabl...
World Cup 2026 Prediction Engine - A probability-first forecasting system that predicts three-way international match outcomes from leakage-safe Elo, FIFA ranking, form, rest, venue, and competition features, then propagates that uncertainty through the official 48-team World Cup structure. Across 384 chronologically held-out World...
England vs France: Third-Place Match Analysis - A reproducible, multi-source analysis of England's 6–4 win over France in the 2026 World Cup bronze final. The study tests whether the match behaved like a low-pressure exhibition, then separates physical effort, defensive activity, control outcomes, score state, and finishing variance. Its central...
Argentina Comeback Analysis: How Rare Was 3–2 vs Egypt? - A statistical and counterfactual investigation of Argentina's late 3–2 comeback against Egypt in the 2026 World Cup Round of 16. At each tested late-match checkpoint—70, 75, 78, and 80 minutes—no historical World Cup match in the 964-match sample ended in a win from a two-goal deficit; at 70 minute...
InterviewMate AI - Virtual Interviewer Assistant - An interactive AI interviewer capable of holding dynamic voice conversations with candidates in real-time. Features WebRTC with the OpenAI Realtime API for sub-500ms latency, Resume Personalization via edge PDF extraction, a Recruiter Dashboard built on Firebase, and an automated 100-point evaluati...
Love Notes AI - A personalized Valentine's Day message generator powered by Google Gemini API. Features smart generative networking, universal design for any recipient, and a beautiful, responsive UI built with React, Vite, and Tailwind CSS. Create heartfelt, funny, or poetic messages in seconds.
PRD Generator - An enterprise-grade AI-powered web application for Product Managers to automatically generate Product Requirements Documents (PRDs) and export them as professionally formatted PDFs. Features Gemini 2.5 Pro for deep-reasoning PRD creation, BYOK (Bring Your Own Key) API key security, rich Markdown ed...
Financial Wellness Agent (FWA) - Production-ready AI-powered personal finance PWA with multi-agent architecture and real-time market analysis. Features 6 specialized AI agents (Planner, Report Generator, Transaction Parser, Budget Analyzer, Goal Advisor, Market Agent), AI-powered stock analysis for IDX and global markets, portfoli...
Case Vault - AI Detective Visual Novel - A limitless, AI-generated noir detective visual novel where every case, suspect, and clue is uniquely created by Gemini 3.0. Features procedural storytelling, real-time AI interrogation with personality-driven suspects, dynamic visuals via Gemini 3.0 Pro Image Preview, and episodic progression acro...
FocusForge - Learning Focus App - A Tauri 2.0 desktop application for enhanced learning through Pomodoro-style focus sessions, AI-powered knowledge consolidation, and gamification. Features TensorFlow.js face detection for distraction tracking with eye gaze analysis, Gemini AI (2.5 Flash / 3 Pro) for Socratic learning analysis, AI...
Universal Commerce Protocol (UCP) Agent - A production-ready AI Agent implementation of the Universal Commerce Protocol (UCP). Features a federated multi-shop architecture with 3 independent stores (Main, Budget, Luxury) and an intelligent shopping agent powered by Gemini 2.5-Flash. The agent handles decentralized search, real-time invento...
ProcureMind - AI Procurement Assistant - An AI-powered procurement assistant designed to streamline the Request for Quote (RFQ) process. Features AI Parser for raw RFQ emails, Semantic Search for historical product matching, and Google Gemini 2.5 Flash integration for analyzing pricing and drafting professional responses. Built as a wrapp...
Meeting Summarizer - A full-stack web application for transcribing meeting recordings and generating AI-powered summaries. Features OpenAI Whisper for audio transcription, GPT-4o-mini for intelligent summarization, and Clerk for user authentication. Deployed on Google Cloud Run with a complete CI/CD pipeline via GitHub...
Baseline Pro - Tennis Coaching Platform - A mobile-first tennis booking platform with Google SSO, interactive heatmap schedule, and comprehensive admin tools. Features include multi-step booking flow with class type selection, birthday voucher system, skill badge tracking, YouTube tutorial library, and Google Calendar integration. Built fo...
Pandas vs Polars vs DuckDB Benchmark - A comprehensive benchmark of Pandas, Polars, and DuckDB on a 16GB laptop processing 13.1 million rows (Susenas 2024). Demonstrated 98% memory reduction and 100x speedup using modern libraries. Includes detailed performance analysis across filtering, aggregation, and join operations.
Cybersecurity Analyzer Agent - A web-based tool designed to identify security vulnerabilities in Python code. Features AI-Driven Analysis using OpenAI's agents, Static Code Analysis with Semgrep via MCP, and an interactive chat interface. Architected for deployment on serverless container platforms like Azure Container Apps and...
Trader Agent Simulator – Autonomous Portfolio Management - Built an intelligent trader agent using the OpenAI Agents SDK, capable of alternating between trading and rebalancing its portfolio using real-time insights from a researcher agent. Leveraged advanced async context management for clean multi-server handling, and integrated support for multiple LLM...
Galaxy Morphology Classification - An interpretable deep learning project for classifying galaxy morphologies (Elliptical vs Spiral) using the Galaxy Zoo 2 dataset from SDSS images. Achieved 99.07% accuracy and ROC-AUC 0.9995 with ResNet18 transfer learning. Features Grad-CAM and Integrated Gradients for scientific interpretability.
Text2SQL — Fine-tuning a ≤3B LLM for SQL Generation - End-to-end pipeline for fine-tuning a small (≤3B) open-source LLM to convert plain-English questions into executable SQL. Fine-tuned Qwen2.5-Coder-1.5B with QLoRA via Unsloth on the BIRD benchmark, evaluated by execution accuracy on real SQLite databases. Lifted the valid-SQL rate from 40% → 73.5%...
IBM Applied Data Science Capstone - A comprehensive data science project predicting Falcon 9 first-stage landing success. Involves Web Scraping, Data Wrangling, EDA with SQL & Visualization, Interactive Dashboards (Dash), and Machine Learning Classification.
Career Digital Twin (RAG Chatbot) - Built and deployed a personal agent to represent my skills and experience to potential employers, automating the initial stages of job applications.
Indonesian Parliament Activity Chatbot - This project implements a chatbot that can answer questions about the activities and agendas of the Indonesian Parliament members. It utilizes a SQL database containing agenda data and leverages large language models (LLMs) through the Langchain library to interact with the database and provide nat...
Telco Churn Analysis - Developed a churn prediction model using AllKNN with hyperparameter tuning, focused on minimizing false negatives. Achieved 93.7% recall, reducing potential high-risk churn losses by $18.8K and cutting total misclassification costs by $48.5K, outperforming benchmark models like XGBoost and Random F...
Airbnb Data Analysis - Analyzed Airbnb listings in Bangkok to identify peak-season revenue opportunities for December. Implemented dynamic pricing, extended-stay discounts, and last-minute deals. These optimizations increased total December revenue by 7.6%, generating an additional ฿3.9 million in high-demand areas.
Waste Classification with Transfer Learning - A computer-vision project (IBM Deep Learning with Keras & TensorFlow capstone) that classifies waste images into Recyclable and Organic categories using transfer learning with VGG16. Demonstrates an end-to-end deep-learning workflow — image preprocessing with ImageDataGenerator, feature-extraction...