darkmice
English · 简体中文
把 AI 能力做成可运行、可交付、可验证的系统。
I turn AI capabilities into systems that can run, ship, and earn trust.
Projects · Live Demo · Case Studies · Talon · Sandbox · Formily
What I build
| Direction | What it means in practice | |
|---|---|---|
| 01 | AI execution systems | Agent orchestration, local execution, tool protocols, sandboxing, and human-in-the-loop delivery. |
| 02 | AI-native data infrastructure | Multi-model data engines that bring SQL, KV, time series, queues, vectors, full-text search, GEO, graph, and AI into one runtime. |
| 03 | Developer products | From product definition and interaction design to SDKs, component systems, desktop clients, docs, packaging, and release workflows. |
| 04 | Quality systems | Evidence-first engineering: adversarial review, explicit risk boundaries, failure-path testing, and verifiable acceptance. |
Public evidence and disclosure boundaries
Most production work is private. Each label below describes only the linked public artifact; it does not make private production implementation independently verifiable.
| Evidence | What you can inspect |
|---|---|
Evidence Lab → Live Demo |
A browser-only clean-room simulation of task steps, intervention, reconnect behavior, evidence, and a human acceptance gate. |
Engineering case studies → Case Study |
Three bilingual deep dives covering agent delivery, a multi-model data engine, and a recoverable coding sandbox. |
Talon releases → Public Release |
Cross-platform binaries and libraries; SHA-256 checksums verify file integrity, not publisher identity. |
Talon · Talon Sandbox Product Site |
Public product positioning, documentation, download, Playground, and sign-in entry points; not implementation proof. |
Public repositories → Public Code |
Inspectable implementation across data, MCP, structured documents, UI, and quality workflows. |
Selected work
| Project | Focus | Evidence |
|---|---|---|
| Formily | Cross-device, high-performance form solution; verified contributions | Open Source |
| Talon Pilot Studio | Public product surface for task decomposition, execution visibility, previews, and acceptance; private orchestration is excluded | Public Code · Case Study |
| Talon | SQL · KV · TimeSeries · MQ · Vector · FTS · GEO · Graph · AI in one runtime | Product Site · Public Release · Case Study |
| Talon Sandbox | A recoverable computer boundary for coding agents: workspace, terminal, browser, preview, and policy | Product Site · Case Study |
| Talon MCP | 38 bounded MCP tools across all nine Talon engines | Public Code |
| Talon Doc Runtime | 30+ semantic components; compact DSL to interactive deliverables | Open Source · Live Demo |
| Talon UI | Design tokens and a 45-component React product system | Open Source |
| Dark Tribunal | Risk-driven adversarial review connected to implementation and acceptance evidence | Public Code |
My working model
Product intent
↓
System boundary → Protocol & data model → Implementation
↓ ↓
Risk ledger ← Counter-review ← Evidence
↓
Verified delivery
- Product before plumbing — start from the decision or workflow the system must improve.
- One source of truth — make contracts, state ownership, and responsibility boundaries explicit.
- Local-first where it matters — keep execution and data under the user’s control when possible.
- Evidence over optimism — a green test or a pushed commit is not the same as an accepted outcome.
Stack I use to ship
Building at the intersection of AI agents, data systems, product engineering, and delivery quality.