Web platform + AI agents
DolphinTask
From browser-only POC to full serverless production — a workload platform where humans, AI agents, and LINE share one database.
- Client
- Self-initiated product
- Year
- 2026
- Role
- Solo — design, build, infra
- Stack
- Hono, Cloudflare Workers, D1, React, TypeScript, MCP
The problem
Daily workload tracking lived in spreadsheets and paper. Offices needed actual vs standard time per activity, capacity, and utilization — but nobody wanted another heavy system to install and maintain.
What I built
A workload platform that started as a client-side POC (IndexedDB, Excel export by template) and evolved into a production serverless system, one phase at a time.
- 01Workload tracking core — Employees, task master, daily actual vs standard time, capacity and utilization reports, Excel export, JSON backup and safety restore.
- 02Time tracking with geofence — Clock in/out with GPS check against work areas, break sessions, and enforced minimum hours.
- 03AI-native by design — Every capability is an MCP tool — AI agents log work and run reports through the same permission model as humans.
Platform
The whole system runs on Cloudflare: Hono Workers serve the REST API, MCP server, and LINE webhook; D1 holds the 25-table database; Queues absorb long LLM turns; Workers Assets serve the React SPA. Secrets live in Cloudflare Secrets Store. It went from POC to production at dolphin-task.pyinnox.online with no servers to babysit.
Results
Next project
Marina Management System