Hsu Heng-Cheng許恒誠 / AI · BACKEND ENGINEER

I build the systems
that hold it together.

Agent orchestration, retrieval pipelines, and the backend that keeps an LLM demo standing up in production.

cat /info/AudoSt0ck1ng hhc-resume.pdf

Stentor

A local agent workspace for Claude Code and Codex — supervised delegation, with a voice.

One desktop process keeps many durable conversations. The journal lives in SQLite, so closing a window, restarting a provider or switching models never throws the thread away — and a background worker reports back into the conversation that dispatched it.

Project work runs through a supervised pipeline: read-only repository research, a requirements interview, implementation inside an isolated git worktree, then an independent read-only review. FIXABLE findings go back to the original implementer for another round; anything short of an explicit PASS fails closed and stops for you.

research → requirements → worktree → review → pass / fix / ask

And it talks. A herald narrates the pipeline through a speaking avatar and speaks up when a run actually needs a decision — so the waiting in a vibe-coding session stops being time you sit and watch.

Agent Orchestration
Claude Code
Codex
MCP
Git Worktree
Review Loop
Persistent Memory
SQLite
TTS
STT
Python

Private repository

OTHER PROJECTS

Persona Contributor

Cross-platform AI avatar application

VRM
Electron
TypeScript
OAuth
MCP

My contributions

VRoid Hub OAuth integration
MCP performance improvements
User–avatar interaction features

View on GitHub

AVATAR Co-author

A lipsyncing VRM companion for work, study, and everyday listening.

VRM
VRoid Hub
OAuth
MCP
Motion Bus
Caching

My contributions

VRoid Hub OAuth integration
MCP service and motion bus integration
Caching and performance improvements

View on GitHub DOI

KyoKyoBot Author

Chat bot with user-defined characters — RAG over its own lore and chat history

Discord
RAG
LlamaIndex
Qdrant
FastEmbed
Memory
LLM

What it does

Any character you define — isekai, novel cast, or a plain assistant
Per-channel chat history, on local files or a database
Long-term memory shared across channels, grounded by RAG
Many people in one channel, talking to one character

Private repository

Other contributions

// open source

LlamaIndexBug fixes and improvementsView on GitHub text-to-vrmaClaude API supportView on GitHub CCUMVL-Vehicle-ReID DatasetVehicle re-identification dataset, built and released during my master'sView dataset

Experience

// career

2024.11
– 2026.07

MaiAgent Co., Ltd.Senior Backend Engineer

Built core agentic conversation and workflow features for a production AI assistant platform; asynchronous execution cut processing time by over 60% in key use cases.
Redesigned the backend architecture — decoupled assistants from knowledge bases, added multi-source retrieval with QueryFusionRetriever.
Built and maintained a Text-to-SQL pipeline for natural-language database queries.
Integrated third-party communication platforms and implemented MCP OAuth authentication.
Automated engineering workflow with GitHub Issues, Sentry and Slack bots for monitoring and alerting, taking the conversation service error rate from peaks of 5% to consistently under 0.5%.

2022.09
– 2024.08

National Chung Cheng UniversityM.S. in Computer Science and Information Engineering

2018.03
– 2021.02

Ares International CorporationSenior System Designer

Developed and maintained core backend and batch systems.
Maintained QA and CI environments and automated software packaging, cutting packaging time by over 80%.

2013.09
– 2017.06

National Chung Cheng UniversityB.S. in Computer Science and Information Engineering

05 / ALWAYS TIME FOR TEA

Take a seat.

Contact me.

03:00 / Time stays. Ideas wander.

I build the systemsthat hold it together.