Three things: turn model capability into software people actually use, hand repetitive work to automation, and organise scattered information into something searchable. We own the whole chain — requirements, design, implementation, deployment and operations.
Each product gets its own domain and its own site, and every entrance starts here. They solve unrelated problems but share one engineering standard: readable interfaces, explicit boundaries, reversible releases.
Every AI competition, in one catalogue
Turns AI competition notices scattered across dozens of official sites, chat groups and newsletters into one filterable, subscribable, reminder-backed catalogue — so finding a contest stops being manual labour.
Put the prep back in front of the camera
A locally running AI copilot that turns narration, screen demos and Q&A into structured notes, chapters and follow-ups in real time — a deliverable recap the moment the stream ends, with no data leaving the machine.
Run your AI team like you run a company
An operating system for digital employees: roles, responsibilities, task flows and hand-off contracts are made explicit, so multiple agents collaborate by position instead of each answering in isolation.
A bilingual technical guide to an open-source robot duck
A bilingual technical reference for the open-source bipedal robot Microduck: hardware specs, software architecture, the sim-to-real training pipeline and the OTA update flow — fully pre-rendered and readable offline. An independent guide, not an official site.
Information architecture, visual systems and responsive implementation. The stack follows the goal: Vue 3, Astro, or zero-dependency native ES modules with static pre-rendering.
Node services, REST and JSON-RPC contracts, SQLite and PostgreSQL persistence. Contracts get fixed before implementation so nobody has to guess field names during integration.
Local and hosted models wired up with structured output, failure fallbacks and call-level observability. Anything solvable on-device stays on-device.
Playwright browser automation, multi-threaded scheduling, cloud instance orchestration — click-by-click workflows turned into scripts you can re-run.
Multi-source ingestion, de-duplication, incremental updates and retrieval. Agree on the definitions first; otherwise automation only produces bad data faster.
Docker Compose orchestration, a single Nginx ingress, mixed static and dynamic hosting, plus access logs, traffic reporting and event tracking.
What each layer uses today; swapped when a project needs otherwise
| Layer | Choice |
|---|---|
| Front end | Vue 3 · Vite · Astro · native ES modules with static pre-rendering |
| Back end | Node.js 22 · REST / JSON-RPC · SQLite · PostgreSQL |
| AI | Local inference (LM Studio / ONNX Runtime) · structured output · model routing |
| Automation | Playwright · multi-threaded scheduling · cloud service APIs |
| Deployment | Docker Compose · Nginx reverse proxy · read-only static mounts |
| Observability | Access logs · GoAccess traffic reports · first-party event tracking |
Whether you want a new product or want an existing workflow automated, describe the situation directly. Arriving with a concrete problem beats arriving with "we want to use AI".