Skip to content
View kittimzhe's full-sized avatar
  • Beijing City
  • 14:54 (UTC +08:00)

Block or report kittimzhe

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
kittimzhe/README.md

Hi, I'm Mengzhe Gan 👋

Full-Stack Engineer · AI Application Engineering · Agent Tooling
Building observable, auditable, and measurable infrastructure for production-grade AI agents.

I'm a software engineer with hands-on experience in full-stack delivery and AI application engineering. My career started in application software development and has grown into end-to-end engineering ownership across frontend, backend, system integration, and production delivery.

I currently focus on turning AI capabilities into real productivity systems for enterprise workflows — not just impressive demos, but systems that can be integrated, governed, evaluated, and operated over time.

Governance and operations separate production systems from experimental AI demos.
The depth of engineering determines how far AI value can be realized.


🧭 About Me

I have worked on complex business systems as well as AI agent systems built from zero to one. Rather than treating AI applications as isolated model calls, I care about how they enter real workflows, connect with existing systems, respect permission boundaries, and remain understandable after execution.

My current work revolves around combining:

  • LLM service integration and orchestration
  • Knowledge bases and retrieval-augmented generation
  • Tool calling and agent workflows
  • Multi-turn session state management
  • Permission models and security boundaries
  • Session evidence, auditability, and postmortems
  • Agent behavior evaluation and regression tracking
  • Full-stack system integration and delivery

I’m especially interested in the engineering gap between "the agent worked once" and "the agent can be trusted as part of a production workflow."


🔧 What I Build

AI Agent Tooling

Infrastructure around real agent sessions: recall, transcript export, evaluation, and regression tracking — so agent behavior can be searched, reviewed, compared, and improved.

Enterprise AI Applications

AI systems for productivity scenarios such as office workflows, developer efficiency, operations support, and knowledge management, with attention to governance, observability, and operational control.

Full-Stack Business Systems

End-to-end applications covering frontend interaction, backend services, data modeling, permission design, integration, deployment, and delivery documentation.


🧩 Featured Projects

A self-hosted knowledge workbench with a visible multi-agent pipeline.

InsightLoom turns saved articles, papers, links, and notes into actionable knowledge. It uses a visible multi-agent workflow to classify, summarize, link, challenge, and assemble knowledge proposals, while keeping humans in control before anything is written to the knowledge base.

Highlights

  • Provides a self-hosted personal knowledge workspace for articles, notes, papers, and links
  • Uses a visible multi-agent pipeline: classifier, summarizer, linker, challenger, and assembler
  • Keeps knowledge writes human-approved: agents create proposals, users approve before Markdown is persisted
  • Stores knowledge as plain Markdown files, making it Obsidian-compatible and avoiding vendor lock-in
  • Supports semantic search, RAG Q&A, browser capture, approval workflow, and daily knowledge digests

Tech: Python · FastAPI · LangGraph · LLM · RAG · SQLite · JavaScript · HTML/CSS


Evidence-grade session transcript export for DeepSeek Harness.

A DeepSeek Harness plugin that exports agent sessions into human-readable HTML, Markdown, JSON, and archive formats. It focuses on auditability, reproducibility, and operational review.

Highlights

  • Reads canonical session logs through ctx.sessionQuery, avoiding recorder drift
  • Generates self-contained HTML reports with KPI cards, turn timelines, tool rankings, and error highlighting
  • Supports Markdown, JSON, ZIP archives, SHA-256 manifests, and review bundles
  • Provides masking and hash-based redaction for sensitive information
  • Useful for debugging, postmortems, team review, and delivery evidence

Tech: TypeScript · JavaScript · DeepSeek Harness · Plugin System · Session Query


Cross-session full-text recall for DeepSeek Harness agents.

A model-facing recall tool that allows agents to search their own past session transcripts, such as previous decisions, bugs, configurations, or project context.

Highlights

  • Exposes a typed recall tool for model-facing session search
  • Searches original transcript events instead of lossy LLM-generated memory summaries
  • Uses persistent full-text indexing with CJK fallback support
  • Provides explicit scope controls including cwd scoping, time filters, tool filters, and all-project policy
  • Includes redaction and access-control options for safer retrieval

Tech: TypeScript · SQLite FTS5 · Agent Memory · Session Search · DeepSeek Harness


Retrospective evaluation for real DeepSeek Harness agent sessions.

A deterministic evaluation layer that grades sessions that already happened and compares performance across sessions, without benchmark authoring or LLM judges.

Highlights

  • Produces reproducible grade cards from persisted session logs
  • Measures reliability, re-ask signals, and tool-call load
  • Supports /eval, /eval-diff, and /eval-history
  • Helps detect regressions after prompt, model, plugin, or configuration changes
  • Designed for real-world agent operations rather than synthetic benchmarks

Tech: TypeScript · JavaScript · Agent Evaluation · Regression Tracking · DeepSeek Harness


A hands-on guide to building DeepSeek Harness plugins.

A practical authoring guide based on real plugin development experience, covering plugin structure, tool registration, command registration, configuration, debugging, and common pitfalls.

Highlights

  • Built from real experience developing dsh-session-export and dsh-session-recall
  • Covers the practical lifecycle of DeepSeek Harness plugin development
  • Includes bilingual material for both English and Chinese readers
  • Helps developers understand the DSH plugin ecosystem faster

Tech: Documentation · DeepSeek Harness · Plugin Authoring · Developer Tooling


🛠️ Tech Stack

Languages

TypeScript JavaScript Python Java Go C++

AI Engineering

AI Agent RAG Tool Calling LLM

Backend & Tooling

Node.js SQLite npm GitHub Actions

Full-Stack Delivery

React Vue.js Docker PostgreSQL


🧠 Engineering Interests

  • Making AI agent sessions searchable, reviewable, and auditable
  • Turning agent behavior into evidence chains for debugging and postmortems
  • Evaluating real-world agent quality with deterministic metrics
  • Designing governance boundaries for enterprise AI applications
  • Connecting models, tools, knowledge bases, and business systems into stable workflows
  • Building engineering bridges from AI demos to production systems

📌 Current Focus

Agent Memory
Session Evidence
Agent Evaluation
Developer Tooling
AI Application Engineering
Enterprise Productivity
Governance & Observability

📍 Selected Repositories

Repository Focus
insightloom Self-hosted knowledge workbench with a visible multi-agent pipeline
dsh-session-export Evidence-grade transcript export for DeepSeek Harness sessions
dsh-session-recall Cross-session recall and full-text search for agent memory
dsh-session-eval Deterministic retrospective evaluation for real agent sessions

🤝 Connect With Me


🌱 Motto

AI value is not only about generating answers.
It is about entering workflows, connecting systems, and supporting decisions.

Pinned Loading

  1. dsh-session-export dsh-session-export Public

    Human-readable session transcript export for DeepSeek Harness — /transcript writes Markdown/JSON to a host path via ctx.sessionQuery (dsh-plugin)

    TypeScript

  2. dsh-session-recall dsh-session-recall Public

    Cross-session full-text recall for DeepSeek Harness: the model-facing recall tool searches past session transcripts via ctx.sessionQuery, with a persistent FTS index

    TypeScript

  3. build-agent-harness-from-scratch build-agent-harness-from-scratch Public

    🛠️ 从零手写 Agent Harness:从 Prompt 到 Runtime —— 不调框架、用纯 Python + LLM SDK 手写一个能跑的 Agent Runtime

    Python

  4. insightloom insightloom Public

    🧭 洞察织机 InsightLoom — 把收藏变成可用知识:可见 multi-agent 流水线的自托管知识工作台

    Python

  5. dsh-session-eval dsh-session-eval Public

    Retrospective evaluation for DeepSeek Harness agent sessions: deterministic grade cards + cross-session regression diffs over persisted logs — no benchmarks, no LLM judges

    TypeScript