jackie-cqzRESEARCH & ENGINEERING
GitHub

AI SYSTEMS / AUTONOMOUS AGENTS

Intelligence,
beyond the
prompt.

Building AI systems that connect reasoning to action.
Exploring autonomous agents, robotics, and privacy-preserving computing.

Explore my workFind me on GitHub
Student at Beihang University/Interning at Z.ai
SYSTEM THINKINGFIG. 01
Agent system concept mapMemory, planning, tools, perception, and privacy connect to an agent runtime. A feedback loop links action to learning. MEMORYPERCEPTIONPLANNINGPRIVACYTOOLS THE CONNECTIVE LAYERAgentruntime OBSERVE → REASON → ACT → LEARN
CONCEPT MAP · NOT A LIVE RUNTIMEIntelligence is connected.
LLM AGENTS✳VISION–LANGUAGE✳ROBOTICS✳PRIVATE AI

01 / OPEN-SOURCE CONTRIBUTIONS

Building with the community.

Agent memory. Developer tools.

OCEANBASE / POWERCONTEXTContributor

PowerContext

Context that stays with the work.

A context and memory system for work that humans and agents hand off across sessions. It preserves decisions, constraints, progress, and evidence through Memory and Handoff, while Experience and Skills capture reusable approaches. Agents can access that context through the Python SDK, HTTP API, or MCP.

My contribution — ZCode integration

  • CLI & Windows desktop. Added a ZCode plugin that captures prompts as Sources, connects PowerContext MCP tools, and supports project-context workflows.
  • Setup & diagnostics. Added setup zcode and doctor zcode, preserving existing settings and sharing one Server URL across Hook and MCP.
  • Memory & Handoff validation. Covered prompt capture, memory generation, fresh-session recall, and nullable Handoff fields with focused tests and integration guides.
PythonJavaScriptMCPAgent Memory
#1752ZCode CLI & desktop integrationMerged
ESENGINE / DEEPSEEK-REASONIXContributor

DeepSeek-Reasonix

A coding agent for long-running work.

A local coding-agent engine with terminal, desktop, browser, and editor entry points. Configurable model providers, tools, and plugins support different workflows; permission checks, workspace sandboxing, and checkpoints help keep autonomous coding sessions inspectable and recoverable.

My contributions — desktop & CLI

  • Configurable terminal themes. Added persistent Follow app, Light, and Dark preferences, terminal palettes, Appearance controls, and live updates for existing xterm sessions.
  • Correct model grouping. Fixed Desktop and CLI model pickers so base providers stay out of the plugin group while real plugin models remain discoverable.
  • Regression coverage. Added terminal-theme, configuration, and model-catalog tests covering preference changes and the provider/plugin boundary.
GoReactxterm.jsDeveloper Tools
#7210Configurable terminal themesMerged#8097Plugin model-group fixMerged

02 / PERSONAL PROJECTS

A few things I've built.

All repositories
01 — AGENT INFRASTRUCTURE

AgentMeld

A local-first multi-agent workspace for task orchestration, context engineering, and inspectable execution.

TypeScriptNext.jsSQLite
Multi-agent systemsSource code AgentMeld
02 — AGENT TOOLING

DSH × TypeSafe Jev

Typed decision tools for DeepSeek Harness, with configurable guardrails and Web UI result cards.

TypeScriptTyped DecisionsPlugin
dsh-jev-pluginSource code DSH TypeSafe Jev

03 / CURRENT QUESTIONS

Questions I keep coming back to.

Beyond a single response.

[ 01 ]

Memory & context

MEMORY & CONTEXT

How can agents retain experience? Exploring memory architectures for persistent learning and contextual reasoning.

[ 02 ]

Agent collaboration

MULTI-AGENT SYSTEMS

How can agents work together on complex tasks? Exploring collaboration, orchestration, and scalable agent networks.

[ 03 ]

Evolution & adaptation

EVOLUTIONARY AGENTS

Can agents improve how they work through feedback? Exploring frameworks for self-improvement and autonomous adaptation.

04 / A LITTLE CONTEXT

From models.
To the real world.

I'm jackie-cqz, a student at Beihang University, currently interning at Z.ai. I work on AI systems, autonomous agents, and privacy-preserving machine learning.

I've worked on planning and tool calling for LLM agents, VLM-powered robotic manipulation, reinforcement learning for robot control, and fully homomorphic encryption for private AI inference. I'm interested in making models part of reliable, understandable systems.

AGENTSPlanning · Tool calling · Collaboration
ROBOTICSVision-language models · RL control
PRIVATE AIHomomorphic encryption · Private inference
TOOLS I BUILD WITH
PythonTypeScriptCC++

KEEP THE CONVERSATION GOING

Good systems start with a conversation.

More code, experiments, and open-source explorations on GitHub.

GitHub / jackie-cqz