homsy
Early working pilot · deployed and tested end-to-end

How it was built

A trust-first home operations companion.

Homsy helps homeowners understand what needs maintenance, when it should happen, and what to do next. Every important action stays under their control.

01 · Story

Why I built Homsy.

The idea started while reading an article that mentioned a home-maintenance task in passing. It made me realize I had no idea how often that task should happen, or what else around my house I might be missing.

The information existed, but it was scattered across articles, manuals, regulations, supplier recommendations, old emails, and other people's memories. Coordinating individual suppliers for every part of the house was its own separate frustration.

A second motivation ran alongside it: understanding AI-assisted product development by completing the whole loop: idea → research → product definition → UX → architecture → implementation → testing → deployment → users. AI shortened the distance from idea to working prototype; I still owned the product judgment and every final call.

“What if I could describe my house once and receive a practical maintenance plan showing what needs attention and when?”

The first prototype was intentionally small

01

Describe the house and its equipment

02

Generate a maintenance plan

03

Add it to a calendar

04

Receive reminders

Preview of the Homsy story: from a solar-maintenance article to a working, AI-built pilot.

Where this could go next

Talking to other homeowners showed me the friction wasn't only mine. These ideas came out of it. They are hypotheses, and none has shipped.

  • Ready-to-send supplier communication
  • Automatic follow-ups on outstanding requests
  • Recurring service workflows
  • Grouped requests across similar homes
  • Coordination between nearby homeowners
  • Better supplier pricing through aggregated demand
Open the interactive story

Beyond engineering

Name, brand, and a squirrel.

Alongside the code, I explored the name, a mascot and its personality, and the visual language. The mascot made an operational product easier to approach, and the visual work let me explore UX before writing production code.

About the name

My kids sometimes call me ‘veverka’ (squirrel) because I love snacking on nuts. Squirrels store, sort, and keep track of what they've got, which fits a product about keeping track of a house. And “Homsy” leans on “Home.”

WelcomeOnboarding greeting
ThinkingPlan generation
Checklist“Vše v pořádku” status
TechnicianNeeds attention
SuccessEmpty task list
EmailCommunication
ReminderCalendar surfaces
Brand palette

Forest-green “homsy” for trust, lime “.ai” for accents. Status colors avoid red, so even a failed state shows as calm gray.

Brand green#145A37
Lime#78B63E
Green soft#E5EDE6
Lime soft#E9F2D8
Text#1F2A26

02 · Product

How Homsy works, one chimney at a time.

A common Czech example: periodic inspection of an active flue path. Homsy's shipped rule applies a 12-month default for any home with a chimney (§ 43 zákona č. 133/1985 Sb.; vyhláška č. 34/2016 Sb.). That is a reasonable baseline, but check what applies to your own house.

Across every step: you can change recommendations and override intervals. Nothing is sent without review. Communication is always draft-first.

03 · Build

Building it with AI.

AI was central to how Homsy was researched, designed, implemented, tested, and iterated.

In the shipped product, AI does not make live maintenance decisions. Rules do.

ClaudeProblem exploration, maintenance-domain research, framing the trust-first positioning
ChatGPTParallel research and idea stress-testing
ClaudeDrafting the PRD, ARCHITECTURE.md, DESIGN.md and the trust/Speed-1-2-3 model
Markdown specsDated specs for the data model, tech stack, and design audit
AI-assisted conceptingMascot poses, wordmark logic, and the dashboard mockup
Hand-authored SVGVector wordmark and house mark, refined by hand
Claude / Claude CodeRepository, KnowledgeProvider and CommunicationChannel seams, OAuth scope strategy, and every reconciliation (e.g. SQLite → Postgres) documented in the specs
Plan workflow11 dated implementation plans, each scoped and checkpointed before being built
Claude CodePrimary pair-programmer from the first scaffold to a deployed pilot
VS Code / CursorEditor, via a Dev Containers workflow
Next.js 15 · React 19 · TypeScriptFramework and language, strict mode
pnpm · ESLint · Prettier · PlaywrightPackage management, linting, formatting, visual tests
GitHubSource of truth for every change
VercelHosting, preview deploys per branch, Vercel Cron for the daily job
Neon Postgres + DrizzleOperational database and migrations
Typecheck, lint, build checksNo full CI test suite yet (see Current state)
Slack (Incoming Webhook)Ops only: a daily digest of new waitlist signups. Not a homeowner-facing feature.
Vercel Web AnalyticsBasic traffic visibility
Follow-ups logA running log of deferred shortcuts, verified fixes, and reverted experiments

04 · Architecture

Your data stays yours. Homsy just coordinates.

A deterministic knowledge layer produces recommendations. House data lives in your own Google Sheet. Neon holds only operational metadata. Gmail receives drafts only. The calendar is a one-way signed feed.

Homeownerdescribes house, approves actions
Homsycoordinates the workflow
Deterministic knowledge layerrule matching → recommendations
Your Google Sheethouse data, user-owned
Neonsessions and tokens only
Gmail draftnever auto-sent
ICS calendar feedone-way, signed URL
You review & approve before anything leaves Homsy
Supplierreceives only what you send
Human & externalHomsy appDeterministic logicUser-owned dataOperational data

05 · Trust

AI should earn autonomy through trust.

Homsy moves from assistant to supervised operator and later to a more autonomous coordinator. Each step waits until the previous one has proven safe and predictable.

SPEED 1MVP · Working

Reminders & suggestions

  • Recommendations & intervals from the rule engine
  • Reminders via subscribed calendar feed
  • Prepared Gmail drafts, never sent automatically
  • Manual approval on every action

This is the MVP, and it is what exists today.

SPEED 2Future / vision

Supervised automation

  • Approved email templates sent automatically
  • Recurring maintenance scheduling
  • Automatic supplier follow-ups
  • User-defined trusted suppliers & rules

Not built yet, though the architecture leaves room for it.

SPEED 3Future / vision

Trusted autonomy

  • Proactive supplier coordination
  • Recurring operational workflows
  • Full audit logs & configurable permissions
  • Emergency override

A long-term vision. Speeds 1 and 2 have to prove reliable first.

Designed to assist, not silently take control.

Human approvalImportant actions require an explicit click.
Minimum OAuth accessNarrowest scopes, requested only when needed.
ExplainableEvery suggestion shows its why and its source.
User-owned dataHouse data lives in your own Google Sheet.
Deterministic logicRecommendations come from matching maintained rules.
Draft-firstSupplier emails stay drafts until you send them.
Progressive autonomySpeeds 2 and 3 come only after Speed 1 proves reliable.
Operational recoveryErrors surface inline. Nothing retries silently.

Trade-offs, made on purpose.

DECISIONBENEFITACCEPTED TRADE-OFF
Human approval instead of hidden actionBenefitNothing surprises the user; trust compounds.Accepted trade-offSlower to feel “magical”; every win needs a click.
Draft-first instead of automatic sendingBenefitZero risk of a wrong email sent on your behalf.Accepted trade-offMore friction per supplier contact until Speed 2.
Minimum OAuth instead of convenienceBenefitSmaller attack surface; avoids sensitive-scope review for Drive.Accepted trade-offNo inbox reading, so supplier replies aren't tracked yet.
Deterministic rules instead of an LLMBenefitEvery recommendation cites a real legal or manufacturer source.Accepted trade-offCoverage is only as good as the maintained rules.
Lightweight MVP instead of a platformBenefitOne Sheet, one deploy and one Postgres DB, so one person can debug it.Accepted trade-offSingle-house, single-user only for now.

06 · Current state

An early pilot, built for learning.

Homsy today is an early pilot. Here is what works, what is still rough, and what real homeowners need to tell me next.

Working now6
  • Onboarding and home setup: profile, assets, suppliers
  • Rule-based recommendations, 3 strictness levels
  • Task, asset, and supplier management
  • Gmail draft preparation, verified by hand
  • Signed calendar reminders (ICS, rotate/revoke)
  • English and Czech localization
Still rough5
  • Google OAuth is allow-listed, with a 100 test-user cap
  • Dashboard is slow to open; the cause is known but not fixed
  • Limited automated test coverage
  • No user-facing activity or audit view yet
  • Single-user, single-house pilot
Next to validate6
  • Do homeowners understand Homsy without explanation?
  • Are recommendations relevant and trusted?
  • Does Homsy reveal overlooked maintenance?
  • Is a calendar plan valuable?
  • Are prepared supplier drafts valuable?
  • Which workflow should be automated next?

Becoming a real product.

Validating Speed 1 with real homeowners is one track. The other is turning Homsy into something people can find, sign up for, and use.

PHASE 1· Working

Early MVP

  • Core loop: profile → plan → tasks → reminders → Gmail drafts
  • Trust-first: minimal OAuth, user-owned data
  • Single builder, single house, real code
PHASE 2· Happening now

Pilot

  • Real homeowners using it besides me
  • Validating usefulness and trust
  • Public availability, pending Google app verification
  • Closing rough edges: errors, slow dashboard
PHASE 3· Future, the fun part

Live version

  • Real local supplier directory and relationships
  • Marketing: messaging and channels
  • The Homsy brand beyond the mascot
  • Real growth once sign-up is open

Working hypothesis

“Leaning on AI at every layer (research, specs, code, ops) is what makes it realistic for one person to explore several directions in parallel instead of betting everything on one path.”

What building it taught me
  1. 01AI dramatically reduces the distance from idea to a working implementation.
  2. 02Good context and real constraints matter more than just asking for code.
  3. 03Faster implementation makes product judgment more important, not less.
  4. 04AI output still requires review, testing, and ownership.
  5. 05When the loop gets this short, building becomes addictive, and the hard part is stopping to validate.
  6. 06The next important learning has to come from real homeowners.
Homsy mascot as a technician

07 · Builder

About the builder.

I'm an engineering leader and hands-on product builder with 15+ years in software engineering and 10+ years leading engineering and technical teams.

Homsy brought me back into the complete product-building loop, from identifying the problem and designing the experience to architecture, implementation, deployment, and preparation for real users.

linkedin.com/in/vladimir-pindura

Homsy is ready enough to stop being only my experiment and start learning from homeowners.

The next round of learning has to come from people who aren't me.