I build agent systems that survive contact with real workflows: operable, auditable, recoverable, and bounded by explicit permissions.

Now

• Building Tenon AI: enterprise digital and AI transformation, plus seven products

AgentHub The control platform for enterprise agent clusters Rebuilt from scratch: one instance per server, a model-gateway key per agent tied to its lifecycle, controlled delegation and multi-agent workflows on a canvas. Live instance · Details	Intent CLI / IntHub Semantic Git for the vibe coding era, and its cloud home Intent, Snap and Decision keep goals, verified progress and durable decisions next to your code; a CLI, an Agent Skill and project hooks give each new session a place to start. Repository · IntHub
Caret An ultralight coding agent in your browser Three tools: read, write and edit. No install, no account and no application server; changes stay in the editor until you have read the diff. Open Caret · Repository	Jig An AI-native multi-agent workbench Born for Session Engineering. You deal only with the main conversation, and approvals and decisions sit in the approval center, on your phone too. Used inside Tenon AI · Details
WeSaid A secret exchange app built on equal terms Private wish pools, mutual matching, dual acceptance and a transactional reveal make safety part of the interaction rather than a policy page. App Store · Web	PassBy A simple password notebook A native SwiftUI password notebook for iPhone, with a web version for the desktop. App Store · Web

Shared infrastructure: an OIDC + PKCE account service (sign-up, email verification, recovery, social sign-in and cross-product deletion) with an integration spec adopted by every product with sign-in, plus one UI/UX standard. Also on GitHub: ai-cultivation, a framework for growing AI from a temporary helper into an auditable, explicitly authorized personal proxy.

AgentHub Enterprise agent management platform Agent creation, runtime, control and collaboration in one platform, now run by multiple maintainers.	SimReadyGen AI 3D asset SaaS for global users An early internal capability turned into a product within three weeks: 486 external accounts and an 86.3% delivery rate.
SimReadyGen API Internal 3D asset generation service One entry point for colleagues, agents and scripts, running in three environments; 2,249 tasks submitted.	BatchRefresh Production batch reprocessing platform Manual scripts replaced by validation, approval, batch rerun and audit; up to 900,000 records in one batch.
Sample Platform Sample showcase and customer delivery Customers preview, request and collect samples in one place; all 105 automated tests pass, in internal beta.

_{Five systems I led during an internship (May 2026 – present). Figures are as of October 30, 2026; details are in my profile.}

• Session Engineering, part 2: one session between me and the rest _2026-11-08
• Session Engineering: organizing agent sessions is also managing your own context _2026-11-06
• Writing every interaction into the model _2026-10-31
• Dynamic intelligence: models that change as they run _2026-10-31

_{More at tenon.asia/blog.}

• Capability is only the beginning. Identity, permissions, state, and recovery determine whether an agent can enter a real workflow.
• One truth, explicit boundaries. Ownership and lifecycle state belong on the server; projections and clients must not silently redefine them.
• Human intervention is a product path. HITL work needs queues, claim/release semantics, quality control, escalation, and a route back into automation.
• A release should be reproducible. I prefer commit-exact builds, immutable artifacts, candidate slots, public acceptance, and rollback over production-side improvisation.
• Safety can be interaction design. The strongest guardrail is often a mechanism users can understand, not another paragraph in a policy page.

• M.Sc. Signal Processing and Machine Learning, Nanyang Technological University · Jan 2026 – Jan 2027
• B.Eng. Electrical Engineering, Shanghai Jiao Tong University · 2020 – 2024

_{The banner, activity card and writing list are rebuilt daily by GitHub Actions from public data.}