From computer vision pipelines to RAG chatbots and MLOps infrastructure, built to run in production rather than sit in a notebook.
The problems, the architecture, and the production reasoning behind each system.
Production financial system for a health-worker cooperative. Double-entry accounting, automated batch processing, five user roles, and legacy Excel migration, deployed on-premise with Docker.
End-to-end ML pipeline for BUY/SELL signals on 45 IDXBLUE stocks. MLflow tracking, Grafana monitoring, Docker deployment, and automated retraining.
Detect-then-classify architecture for high-accuracy object recognition. YOLO/SSD detection paired with custom PyTorch classifiers, optimized for edge deployment.
Retrieval-Augmented Generation chatbot answering questions about my CV. LangChain LCEL + FAISS + FastEmbed (ONNX) + Groq Llama 3.3, evaluated with RAGAS metrics.
A RAG system built from collaborating agents: a router, a retrieval agent, an evaluator, and a summarizer, held together by typed contracts. Works on any document set, from books and manuals to internal knowledge bases.
Playwright-based scraper that automatically searches for 2,500+ product images across 4 Indonesian e-commerce platforms, with anti-bot evasion, scheduled runs, and resume support.
AI-driven chess variant exploring state-space search, evaluation heuristics, and game design patterns, using Minimax with Alpha-Beta pruning.
The gap between a Jupyter notebook and a production system is where most ML projects fail. I build systems with monitoring, evaluation pipelines, and clean architecture from day one.
Whether it's a cooperative's accounting software or a stock prediction pipeline, the standard is the same: ship it, measure it, improve it.
Read my full approach →A live demonstration of Retrieval-Augmented Generation. Ask anything about my background, skills, and projects. Built with LangChain, FAISS, and Streamlit.
Technical deep-dives on ML engineering, MLOps, and building in production.
RAGAS evaluation, retrieval metrics, and why "it works on my machine" isn't enough for production LLM systems.
Why splitting detection and classification into two stages gives you better accuracy, maintainability, and deployment flexibility.
Running MLflow, Grafana, and Docker on modest hardware. Pragmatic MLOps for solo engineers and small teams.