TIM SCHROEDER
Work / Property Station
Homeowner SaaS

PropertyStation

A private home record for homeowners that builds a maintenance plan from the systems in the house and reminds you before each task is due.

Live productFeb–Oct 2026
Property Station homepage with the headline 'Your home, finally on a schedule' beside a laptop and phone showing the home dashboardReal screenshot · public homepage

The problem

A home's manuals, warranties, receipts and repair history end up in drawers, inboxes and phone photos. Maintenance gets missed until something breaks.

Who it's for

Homeowners, and home builders who want to hand each buyer a complete record of the new house.

Where AI is used

Claude answers questions from the home's own record and guides repairs with safety first. Smaller models read documents, receipts and equipment labels, and an image model draws redesign ideas on your own room photos.

Built with

  • Next.js 16
  • Supabase Postgres
  • Row-level security
  • Claude Sonnet 5
  • Claude Haiku 4.5
  • gpt-image-2
  • Stripe
  • Vercel
01 · The problem

A house has no memory

Every home has dozens of systems, each with its own manual, warranty, filter size and service schedule. Most owners keep none of it in one place.

When something fails, the owner hunts for the warranty, the receipt or the last contractor. When they sell, the buyer gets a stack of paper or nothing at all.

  • Maintenance tasks are missed because nothing reminds you before they are due.
  • Documents are not connected to the things they belong to, so finding the right one is slow.
  • Generic AI chat does not know your house and can give unsafe advice in an emergency.
  • Builders hand over a closing binder that the buyer rarely opens again.

Why custom software

Off-the-shelf note and file apps store documents but do not know what a water heater is, when its filter is due, or which warranty covers it. Property Station models the home itself, so plans, reminders and AI answers all come from the same record.

02 · The solution

One private record that runs the house on a schedule

You add your home and its systems. Property Station builds the care plan, links every document to what it belongs to, and answers questions from that record.

A maintenance plan built from your systems

The plan comes from the systems you add, not a generic checklist. One daily job sends reminders before tasks are due and checks local weather alerts.

AI

Ask My Home and the Repair Assistant

Claude Sonnet 5 answers from your own record and says when the record does not know. For emergencies like a gas smell, a fixed safety card written in code appears above any AI answer.

AI

Scan documents, receipts and labels

Claude Haiku 4.5 reads a receipt, a document or an equipment label and fills in the form. You check the fields before anything is saved.

Documents linked to what they belong to

One permit can sit on both a project and the electrical panel it touched. Warranties, manuals and service records stay connected to the right system or room.

AI

Design AI on your own rooms

Sonnet 5 suggests three design ideas from a photo of your room. OpenAI gpt-image-2 draws the 'after' picture.

Builder handoff and Home Snapshot

A builder can preload a new home's systems, finishes and documents and hand the record to the buyer. When you sell, a Home Snapshot shows the next owner what was done to the house.

03 · My role

What I did. What the AI did.

I owned the product and its operation. Claude Code, an AI coding agent, wrote all of the code.

Tim · human judgement

I owned

  • Set the concept, the feature scope and who it is for: homeowners first, with a builder portal for handoff.
  • Made the product and UX calls, recorded in the project's decisions log: the dashboard brief, the homepage mockup, image priority rules, and linking one document to many records.
  • Asked whether the cheaper Haiku model was good enough, and commissioned the blind model test that moved the home chat to Sonnet 5.
  • Commissioned and scoped the audits: launch simulation, security and privacy, accessibility, usability, AI quality and final readiness. I approved the AI evaluation budget.
  • Set the bar for done: a change counts only when a test, a real build or a live check proves it, not when the code looks right.
  • Ran the launch: I unblocked the domain and email setup that needed my accounts, the site opened to the public on 2 Sep 2026, and I am the named owner for alerts.
Claude Code · AI implementation

AI did

  • Wrote all of the application code, the 121 database migrations and the tests.
  • Ran the audits I asked for and wrote the findings, with severity and evidence.
  • Drafted each fix as a pull request with an explanation, then checked it with lint, type checks, a production build and the test suite.
  • Built the AI evaluation scripts and ran the blind model comparison and the golden-set tests.
  • Wrote the runbooks for incidents, backups, cron failures, AI spend and billing.

I decide what to build, what good looks like and what must be proven. The agent writes the code and the evidence, and I review the result before it merges.

04 · How I built it with AI

From idea to production.

The git history covers February to October 2026. Most of the work happened in August and September, in short cycles of build, audit and fix.

  1. Feb 2026 · Foundation

    Existing app put under version control and secured

    The first commit captured the existing codebase. The next added ownership checks to every server action and protected the private routes.

    commits 32e4f33, 78aa857 (17 Feb 2026)
  2. Jul 2026 · Builder portal

    Builder portal and first readiness audit

    Builders got organisations, model homes, clone-to-home and an invite-and-claim handoff. A full production-readiness audit then listed what blocked launch.

    commit e2b7fef (7 Jul) · PR #1 (27 Jul)
  3. Aug 2026 · Hardening

    AI cost limits, launch simulation, tenant isolation

    AI quotas and a spend breaker came first. A simulated launch then found three blockers, and a re-audit found two cross-tenant data holes. All were fixed and pushed to production.

    PR #4 (1 Aug) · PR #24 (21 Aug) · PR #33 (23 Aug)
  4. Sep 2026 · Public launch

    Site opens, operations get eyes

    The site went public with a platform admin console. A dead scheduled-job problem was found and fixed, and uptime checks and branch protection were added.

    PR #45 (2 Sep) · PR #46 (3 Sep) · PR #50 (4 Sep)
  5. Sep 2026 · Audit series

    Security, accessibility, usability and AI quality

    Audits ran back to back. The security audit closed one P0 and five P1 findings. The AI audit added a permanent golden test set.

    PR #220 (23 Sep) · PRs #345–#347 · PR #350 (29 Sep)
  6. Oct 2026 · Launch gate

    Final readiness audit and stricter headers

    A final audit checked deploy, rollback, backups, AI cost limits and support. Payment events now alert the operator, and the full Content-Security-Policy is enforced.

    PR #363 (1 Oct) · PR #364 (1 Oct) · PR #371 (15 Oct)
05 · Technologies

What it runs on, and why.

Every item here is in the project's code or configuration. I chose and connected them; AI coding agents wrote the code that uses them.

AI models

Claude Sonnet 5
Answers home questions, guides repairs, suggests design ideas and reads inspection reports
Claude Haiku 4.5
Reads documents, receipts, paint and equipment labels, and sorts inbound email
OpenAI gpt-image-2
Draws the 'after' picture for a room redesign
Vercel AI SDK
One interface for streaming chat and structured output

App

Next.js 16 and React 19
The web app, server actions and API routes
zod
Checks every form and every AI response against a schema
Tailwind and shadcn/ui
The interface components

Data

Supabase Postgres
The home record, with row-level security on every table
Supabase Auth and Storage
Sign-in, and private storage for documents and photos
Cloudflare R2
A nightly copy of stored files

Services

Stripe
Checkout and subscription webhooks for paid plans
Resend
Reminder and sign-in email, plus inbound email to a home's inbox
Vercel
Hosting and the one daily scheduled job
GitHub Actions
Tests, uptime checks, migration drift checks and file backups
06 · System architecture

How the pieces fit.

A Next.js web app talks to a Supabase database where each owner can only reach their own homes, with AI services called only from the server.

Property Station architectureA Next.js web app talks to a Supabase database where each owner can only reach their own homes, with AI services called only from the server.requestschat, scansroom redesignreads and writesuploadspayment eventsnightly copydue tasksremindersWeb appHomeowners and buildersServer actions and APINext.js 16 on VercelClaudeSonnet 5 · Haiku 4.5Image modelgpt-image-2StripeCheckout and webhooksPostgresSupabase · RLS on every tableFile storageSupabase StorageBackup copyCloudflare R2, nightlyEmailResendDaily jobReminders, weather, notices
InterfaceApp / serverAI serviceData / storageExternal serviceScheduled job
Technical detail for engineers
  • Every public table has row-level security. Content is scoped through the property it belongs to, so one owner cannot read another owner's rows.
  • The privileged database key is used only in server code. The September security audit reviewed all 157 places that use it.
  • Every AI call goes through quotas and a usage ledger, with a global kill switch, per-feature kill switches and a deadline on each call.
  • Photos are downloaded and downscaled on the server before they go to the model, so the AI provider never fetches a private file URL.
  • One daily Vercel cron runs six jobs in order under a shared time budget. Each run is written to a cron_runs table so a missing run is visible.
  • CI runs lint, type checks, a build with no secrets, unit tests, Playwright end-to-end and accessibility tests against a seeded local database, and 13 SQL checks on access rules.
07 · Challenges

Problems worth telling you about.

Real incidents from the project's history, each with how it was found, fixed and checked.

CHALLENGE 01

Scheduled jobs that silently never ran

The reminder, notification and weather jobs had stopped running for at least a week. No error appeared anywhere, and homeowners simply stopped getting reminders.

VERIFIEDA test now fails the build if vercel.json declares more than one cron. The admin console marks a job stale after 36 hours.
Why it was hard, what I decided, how AI helped
Why it was hard
Vercel's Hobby plan caps how many cron jobs a project may register. The project declared three, and Vercel registered none of them without saying so. An earlier fix had addressed the frequency limit but not the count limit.
My approach
An audit I had run on the new admin console found the gap: zero cron calls in the production logs, and the newest notification was seven days old. I wanted a fix that could not come back when a plan or limit changed.
How AI helped
Claude Code collapsed the three schedules into one daily route that runs each job in turn, recorded every run in a cron_runs table, and showed scheduled-job health on the admin console.
The decision
One schedule, not two, because one is inside every plan's limit. A job that has never run had to show up as a row, not be missing, because missing reads as 'no such job'.
Evidence
PR #46 · commit eeec476 · migration 042 · tests/production-readiness/cron-routes.test.ts
CHALLENGE 02

Two cross-tenant holes in the access rules

A re-audit found two P0 holes. A homeowner could change and then read builder-internal rows on a home they had claimed, and every signed-in user could read every other user's private contractor contacts.

VERIFIEDThe SQL regression tests run in CI. The migration was pushed to the hosted database and checked against the live schema.
Why it was hard, what I decided, how AI helped
Why it was hard
Each rule looked right on its own. The read rules had been tightened, but older update and delete rules still matched the same rows. The contractor table was treated like a shared directory when it actually held private, user-created data.
My approach
I commissioned a full production-readiness re-audit before opening the site to the public.
How AI helped
Claude Code wrote migration 032 to gate the update and delete rules, scope contractor contacts to the user who created them, and backfill ownership. It added SQL regression tests to CI.
The decision
Fix the rules in the database, not only in app code, so every path into the data is covered.
Evidence
PR #33 · commit c083d7d · migration 032 · execution/test-visibility-writes.sql
CHALLENGE 03

A launch simulation that caught three blockers

Walking the app as a new user found that sign-out and uploads could hang forever, that every structured AI request failed, and that paid AI output was never saved.

VERIFIEDFixes merged and deployed, migration 030 applied to production, and the app's AI output schemas were each checked to return valid output.
Why it was hard, what I decided, how AI helped
Why it was hard
None of it showed as an error. A safety timeout masked the sign-in deadlock. The insert failures were ignored, so the screen looked fine while nothing was stored.
My approach
I had the whole launch simulated on a local copy of the stack, as a real user would use it, before the site opened.
How AI helped
Claude Code traced the hang to a database call made inside a sign-in event handler, stripped the schema keywords the AI provider rejects while keeping full validation of the response, and added the missing write rules.
The decision
Treat lost paid AI output as a P0. A user who is charged for an answer must get the answer saved.
Evidence
PR #24 · commit 712b46a · migration 030
08 · Results

What is true today.

Property Station is live and public. It went through a series of audits from July to October, and the key fixes are guarded by tests.

452
commits, Feb–Oct 2026
Source: git log origin/main, 2 Nov 2026
353
merged pull requests
Source: gh pr list --state merged, 2 Nov 2026
121
database migrations
Source: supabase/migrations folder count
259
test files, incl. 18 Playwright specs
Source: test and spec file count on main, 2 Nov 2026
60 / 62
AI golden-set runs passed after fixes (40 / 61 before)
Source: AI quality audit, PR #350
21 of 25
blind comparisons won by Sonnet 5 over Haiku 4.5
Source: docs/production-readiness/2026-09-24-haiku-vs-sonnet.md
  • Live at propertystation.com and open to the public since 2 Sep 2026.
  • Ask My Home now leads with safety: the gas-leak case that failed in the AI audit passes, and the safety cases pass 11 of 11.
  • A test fails the build if more than one cron is declared, so the silent-cron failure cannot return.
  • 13 SQL checks on access rules, grants and storage policies run in CI on every database change.
  • The September security audit closed one P0 and five P1 findings, and a production check found no sign they were ever used.
09 · Lessons

What this project taught me.

  1. Silence is not health. A job that never runs throws no error, so record every run and list the jobs you expect.
  2. Test the platform you deploy to. The local database did not reproduce hosted permissions, and one CI check passed while the same rule was false in production.
  3. Compare models blind, on real data. Sonnet 5 won 21 of 25 at about three times the cost per call, and the test also found an error that neither model caused.
  4. Do not leave safety to the model. The emergency safety card is written in code, so it appears even if the model ignores its instructions.
10 · Demo

See the public site

The signed-in app needs an account, so this tour uses screenshots of the public pages.

Guided tour · real screenshots, not a live app
  1. Property Station homepage with the headline 'Your home, finally on a schedule'
    STEP 1 OF 5

    The promise

    The homepage leads with the plan and the reminders, and ends with the Home Snapshot for selling.

  2. Features page showing the app's home dashboard with Home Care, Systems, Rooms, Records and Home Snapshot tiles
    STEP 2 OF 5

    The home record

    Every part of the house has a section: care, systems, rooms, supplies, records, shutoffs and the Home Assistant.

    WHERE AI IS INVOLVEDHome Assistant answers from this record, not from general web knowledge.
  3. How It Works page with three steps: add your home, organize everything, stay on top of it
    STEP 3 OF 5

    Three steps

    Add the home, organize documents against the systems they belong to, then get reminders before tasks are due.

  4. For Builders page headed 'A better home handoff starts here' beside a photo of a new farmhouse-style home
    STEP 4 OF 5

    Builder handoff

    Builders can preload a new home's systems, warranties and documents and hand the record to the buyer.

  5. Security page headed 'Your home information should stay private' with four cards on privacy, account separation, encryption and sharing
    STEP 5 OF 5

    Privacy first

    The security page explains in plain words how each account is kept separate and who can see what.

Let's talk about
what you need built.

Open to AI roles: full-time, contract or consulting. I build AI products, internal tools and automations with AI coding agents.