Jay (CheHao) Liu
Engineering notes
How I think about my own projects: the decisions and architecture behind them, not just what they do.
-
Answers that cite their source
Turning a business’s PDFs into a support chatbot for its website that cites the page behind every answer, with queue-driven document processing, a human review step and an answer cache in front of the model.
-
Letting agents into the SDLC, safely
Plenty of tools make an LLM write code; the harder question is whether you can trust the diff enough to merge it. Every task runs in a disposable Docker sandbox and only becomes a PR after it clears tests, a secret scan and an independent review.
-
An agentic knowledge system
Personal notes pile up, go stale and start to contradict each other. Knowledge Compiler runs an agentic workflow that finds related notes by concept and decides what to merge, update or add, then leaves the final call to you.
-
AI Meeting Management Platform
A team-built meeting platform that turns recordings into actionable follow-ups. An LLM analyses the discussion, extracts action items and helps the team move from a raw transcript to work that can actually be tracked.
-
Grounded, private, cloud-native
A conversational assistant for a user's own Strobe posts and photos, combining Bedrock, grounded vector retrieval, explicit permission-controlled tools and asynchronous media enrichment.
