Le Van Hoang — "Henry"

I started in computer vision — content moderation at social-network scale, then medical and identity documents — and moved with the field into document intelligence and, now, agentic LLM systems. The through-line is the unglamorous part: making a model that works in a notebook survive real users, real latency budgets, and real on-premise constraints.

At Torilab I build multi-persona conversational agents — the interesting problem there is memory, not generation. Getting an agent to remember the right things across sessions, and forget the rest, is what separates a demo from something people come back to.

Before that, four years at Samsung SDS R&D Center building document AI: a legal document platform running a fully on-premise RAG stack with a self-hosted LLM, and an IDP pipeline that digitalized records for a top-3 Vietnamese bank. I also owned the MLOps side — Kubeflow on Kubernetes, which cut deployment time in half across 10+ production models.

I'm currently building LessonAI, which turns a teacher's raw vocabulary list into a standalone interactive English lesson. Open to Vietnamese-language AI and LLM product work.