A LangChain project exploring advanced agent architecture with tool integration, custom hooks, dynamic model selection, multi-model coordination, and middleware implementation. It's designed to showcase best practices for building sophisticated, scalable agent systems.

Users upload a PDF and the system extracts text, images, tables, and formulas. Text is chunked using semantic chunking, stored in PostgreSQL, and embedded with a Sentence Transformers model into pgvector. A local LLM (via Ollama, for full privacy) answers questions grounded in the document — returning relevant chunks, reference images, and metadata alongside each response.

GoFintel is a UN Counter-Terrorism Centre platform used by Financial Intelligence Units to screen transactions, investigate suspicious financial activity, and uncover networks — all while upholding human rights standards. As the lead full-stack developer I owned Flask APIs, PostgreSQL and JanusGraph data layers, Apache Airflow ingestion pipelines, and the React + D3.js investigative visualization layer used by analysts to explore complex connected data.

A mobile-first Progressive Web App serving migrant workers across Thailand, Myanmar, Malaysia, Cambodia, and Nepal. Features include job search, recruitment agency validation, job applications, a CV builder, migrant worker news, polls, and a social tab. Built for thousands of real users navigating labor markets where unvetted agencies pose serious risks.

A full-stack AI web application where construction workers photograph a site and the system determines whether it meets safety standards — using models trained specifically on construction safety data. I owned both backend and frontend, built the APIs consumed by the mobile app, and integrated the AI inference layer into the platform.