I build production AI systems across very different domains — investment research, financial document extraction, ad operations. They share one engineering pattern: typed tool boundaries between model reasoning and any action with consequences, evidence-anchored outputs, human-in-the-loop on anything irreversible, full audit trail.
The model is allowed to be wrong. The system is not allowed to be confidently wrong.
Live: hkfilings.app · hecang.app · asapilot.com
Available to embed with a team taking AI from demo to production — Forward-Deployed / Applied AI Engineer. (LinkedIn)

