← All case studies

Full-stack · self-hosted

Homebase

Our household's own Todoist, built test-first with a local AI model writing the screens.

Role: Product, architecture and an AI-assisted build pipeline
Stack: Kotlin · Ktor · PostgreSQL · React · TypeScript · Tailwind
Status: In daily use by our household

Screens

The real app, with made-up data and placeholder names.

Homebase Today view on desktop with overdue, today and done counts, family filters, tonight's meal and tasks grouped by list — Today on desktop: counts, filters, tonight's meal.

Homebase Today view on a phone with a bottom tab bar and a new task button — The same view on a phone.

A Kitchen refresh project with Plan, Buy and Do sections, sub-tasks, labels and assignees — A project with sections and sub-tasks.

Medications screen with a running-low banner and morning, bedtime and as-needed doses — Medications: doses by time of day, refill warning.

The problem

Our family ran on a commercial to-do app. It couldn't feed our kitchen wall display, didn't know about meals or medications, and kept our lists on someone else's server.

I wanted one place for tasks, shopping, chores, the meal plan and medication doses, shared by the household and shown on the wall, and a way to build it quickly without giving up on quality.

What I built

A full task managerLists in folders, sections, labels, saved filters, subtasks, repeats, reminders, comments, Today, Upcoming, Inbox, search and live updates.

A drop-in for TodoistA Todoist import and a Todoist-compatible API, so our existing lists and habits moved over instead of starting from scratch.

Meals and medicationsA meal library and weekly plan, freezer stock, groceries generated from the plan, and medication doses with history, refills and alerts.

On the kitchen wallPlugins for the Home Screens wall display show tasks, meals and medication doses, through a separate wall API.

How it works

The app itself is a straightforward three-tier system. What is unusual is how the screens get built: each one is a small step whose test is written first, and a local AI model writes the code until that test passes.

The app

Phones & laptopsReact 19 · TypeScript · Tailwind

Material-style shell
Live updates

HTTPS

Ktor APIKotlin · Exposed

Households, lists, tasks
Meals and medications
Todoist-compatible API

SQL

PostgreSQLDocker on a home server

Daily backups, kept 14 days

How each screen gets built

Step + testthe spec

Proven against a throwaway implementation first

step.sh

Local modelAider + qwen3-coder

Writes the screen
Cannot edit the test

runs

check.shtypes + full test suite

Pass or fail, nothing in between

passes

Reviewme

Look at it, mark it done

Key decisions

The test is the specEvery step ships with its test, proven against a working implementation before it is handed over. The model may never edit the test, so a passing step means the screen really does what was asked.
Mistakes become rulesEach new kind of mistake the model makes becomes one line in the conventions file, so it does not make it again.
Compatible on day oneSpeaking the Todoist API meant imports, scripts and habits carried over, instead of asking the family to start from scratch.
Walls get their own keysEach wall display authenticates with its own key against a separate wall API, not with a family member's login.
Boring to runOne Docker Compose file on a home server, a health check, and daily backups kept for two weeks, with restore steps written down.

Where it stands

Today

In daily use by our household on phones, laptops and the wall. The server has 66 tests across 14 files.

More wall plugins and the remaining screens, one tested step at a time.

KotlinKtor 3ExposedPostgreSQLReact 19TypeScriptTailwind 4ViteVitestDockerAiderLocal LLM