> whoami
name: Tianxiang Zhan
role: Graduate
affiliation: University of Electronic Science and Technology of China (UESTC)
location: Chengdu, China 🇨🇳
status: "Graduate from UESTC ↓"
former: M.Eng. @ UESTC (2023–2026) · B.Eng. @ SWU (2019–2023)
mission: >-
Build trustworthy autonomous agents that can reason over, forecast, and act on
complex temporal evidence — with explicit uncertainty and stop-discipline.🔭 Recently graduated from UESTC, pushing agentic time-series systems and uncertainty-aware decision making further.
📡 Latest commits on aion (the harness)
9ce5b6b— feat: add persistent Frontier recovery for premature Goal closure (ztxtech, 2026-10-13)0cffced— fix: resume AION after premature goal closure (ztxtech, 2026-10-13)1e202be— feat: persist AION task state and resume session loops (ztxtech, 2026-10-13)ed8fc5c— fix: scope AION hooks to AION agents (ztxtech, 2026-10-13)270a762— fix: prioritize active tasks on AION board (ztxtech, 2026-10-13)1709cbf— feat: adopt OpenCode V2 plugin on master (ztxtech, 2026-10-12)b0ca149— fix: align gap report example with threshold (ztxtech, 2026-10-12)e42e33c— fix: align goal and self-purify contracts (ztxtech, 2026-10-12)
AION is my flagship project: a time-series harness — a control layer that connects task specification, runtime execution, and result assessment into one stable process for next-generation time-series workloads. Built on top of OpenCode, with 4 layers (Task / Workspace / Execution / Review), 6 agents, 8 protocols, and 17 skills.
AGENT × TIME × UNCERTAINTY— research across three orthogonal axes that intersect in AION: build agents that operate on temporal evidence, and ground every claim in calibrated uncertainty.
🪦 Retired from the README: Django, Flask, MATLAB, Photoshop, scikit-learn, seaborn, pandas, MySQL — the academic-era toolkit. They've been folded into the projects above (or quietly archived).
🛰️ This README is kept in sync with my homepage via self-updating GitHub Actions — see .github/workflows (metrics · aion commits · snake · daily quote).
💬 Time series are 80% assumptions and 20% numerics.
Last refreshed: 2026-10-16 (Asia/Shanghai)

