Projects

Production AI case studies

Each project is documented by challenge, architecture choices, and measurable result.

LLM Platform

Full-Stack Conversational AI Platform

Full-stack conversational AI (Aug 2025 – Oct 2025) using Next.js and Supabase to automate user interactions via third-party messaging APIs.

Automated multi-turn engagement with context-aware RAG responses

  • Next.js
  • Supabase
  • PostgreSQL
  • RAG
  • OpenAI API

Challenge: Product teams needed reliable automated conversations with accurate, context-aware replies across external messaging channels.

Architecture: Built a Next.js + Supabase stack with a RAG conversational engine, real-time sync between PostgreSQL and external data sources, and an admin dashboard for monitoring, human-in-the-loop intervention, and knowledge-base management.

Result: Delivered end-to-end automation for complex multi-turn dialogues with consistent data for reporting and continuous performance tuning.

RAG System

LegCo Insight Bot

AI conversational system (Sep 2025 – Oct 2025) with a RAG pipeline optimized for domain-specific accuracy on legal and policy documents.

Domain-tuned legal Q&A via hybrid RAG on Legislative Council texts

  • RAG
  • Hybrid Search
  • pgvector
  • Supabase
  • OpenAI API

Challenge: General-purpose LLMs lacked reliable accuracy on Hong Kong Legislative Council and related legal corpora.

Architecture: Engineered a RAG pipeline with hybrid search and pgvector on Supabase, a production REST API for the LLM app, and a serverless ingestion pipeline for training data.

Result: Improved domain-specific response quality and enabled seamless integration into production workflows.

Applied AI

Full-Stack ERP / WMS Migration

Client ERP modernization (Aug 2025 – Oct 2025): migrated from Odoo to a custom WMS with React, Node.js, TypeScript, PostgreSQL, and Docker.

Replaced legacy Odoo with a validated, automation-first warehouse workflow

  • React
  • Node.js
  • TypeScript
  • PostgreSQL
  • Docker

Challenge: Legacy Odoo workflows caused manual bottlenecks, data-entry errors, and slow order fulfillment for warehouse operations.

Architecture: Designed an automation-centric flow from order intake to dispatch, a centralized validated data pipeline, and GenAI-assisted spec coding and refactoring during migration.

Result: Reduced manual processing overhead and improved data integrity and order fulfillment accuracy.