Gerald Koh

Education

Aug 2022 – Dec 2025

Bachelor of Science (Computing & Law), Second Major in Artificial Intelligence

Mar 2026 – Present

AI Engineer

Productionised a multi-agent HR assistant using hierarchical LangGraph and LangChain orchestration, dynamic tool routing, task decomposition and self-correction.
Engineered an event-driven agent backbone on Azure using Container Apps, Storage Queues, KEDA, API Management, Front Door, Application Gateway with WAF, managed identities and private networking.
Built SharePoint document-ingestion and retrieval pipelines with Azure Document Intelligence, Azure AI Search, Azure OpenAI, Application Insights and distributed tracing.

Feb 2025 – Oct 2025

Software Engineer, Product

Owned the product roadmap and shipped CaseLens using React, FastAPI and PostgreSQL (Supabase), automating the extraction of unstructured case data into structured, presentation-ready legal analysis templates.
Engineered multi-jurisdictional reasoning pipelines (Singapore, India, UK) for multiple case types through systematic prompt engineering and post-output schema validation, significantly reducing average case analysis review time.

Jun 2025 – Aug 2025

IT (AI Engineering) Intern

Built a modular RAG benchmarking framework evaluating multiple pipeline configurations over unstructured corpora, using metrics such as ROUGE, BLEU, precision and custom domain-specific benchmarks, with vLLM for fast inference and Qdrant vector database, enabling reproducible experimentation and evidence-based configuration selection.

May 2024 – Aug 2024

Data and Digital (Software Engineering) Intern

Developed an end-to-end search interface using HTML, HTMX, FastAPI and LangChain, enabling lexical, semantic and hybrid search types and utilising query expansion for improved knowledge discovery.
Developed scalable, modular web scraping infrastructure using object-oriented principles in Python to support flexible data extraction across diverse sources.
Performed in-depth legacy code audits and refactored critical components, achieving a 90% (~11 hours) reduction in runtime through algorithmic and architectural optimisations.

May 2023 – Dec 2024

Research Assistant (Artificial Intelligence) • Applied AI research for law

Conducted literature reviews on retrieval, summarisation and evaluation, and translated them into iterative agentic pipelines using methods such as Stanford's STORM and programmatic prompting with DSPy, to generate citation-grounded synthetic legal texts.
Built lexical, semantic and hybrid retrieval infrastructure on Vespa with 8000+ court judgments to support RAG-based legal reasoning workflows.
Designed structured query logging and human feedback pipelines to curate training samples for fine-tuning a ColBERT re-ranker.
Engineered end-to-end ETL pipelines using AWS, Docker, and Airflow to automate ingestion of legal regulations and guidelines.

Sharing practical tools I've built to speed up my own workflows, plus some notes from tinkering with AI and performance.

A native macOS desktop app for Excalidraw using Electron and React, with filesystem-backed persistence, native Finder file associations and multi-canvas management beyond browser localStorage limitations.
Experimental branch (feat/mcp): an in-app Copilot that streams diagrams onto the canvas and narrates what it draws.