01 · Tasks die in chat
This started as a learning project with a thesis attached. The learning part: I wanted real native development — SwiftUI on iOS, Kotlin and Compose on Android. I'd tried Flutter and React Native and wasn't happy with either; if I was going to understand mobile properly, I wanted to learn how each platform actually thinks.
The thesis came from watching how companies actually operate. Almost every business I'd seen runs its work through WhatsApp: the boss assigns something in a chat, everyone nods, and by next week nobody can find it — it's buried under three hundred messages, unsearchable, unowned, forgotten. So I built a kind of WhatsApp where the task doesn't die: chat stays natural, but any message can become a Task with an owner, a due date and a state that refuses to disappear.
02 · One Task, two worlds
The design center is a single Task object living in two very different worlds. Personal mode stays light — family and friends, no scores, no approvals, just things that need doing. Official mode is governed accountability for organizations: required due dates, work queues, priorities, blockers, proof submission, review chains, audit trails and a rolling performance score. Same object, radically different enforcement.
Official tasks run on an eleven-state machine — from Assigned through Queued, In Progress, Blocked, Submitted For Review, all the way to Completed or Reopened — plus a derived 'Ignored' signal for tasks someone saw and quietly sat on. Recurrence was treated as launch-critical rather than post-MVP, because organizations live on recurring work: generation is deterministic and idempotent, and it respects each org's working calendar and holidays.
My favorite decision is about trust: the org admin cannot read your work. Admins configure the organization — members, roles, policies, calendars — but they don't get a panopticon; visibility comes only from holding a real operational role. And that boundary isn't a UI checkbox — it's enforced in Postgres row-level security. Organizational values, written as schema.
03 · A frugal, serious stack
The architecture applies the same cost lens I bring to employers' platforms, except this time it was my own money. Supabase gives me Postgres with row-level security by default, auth and realtime — sensitive mutations go through server-side RPCs with audit trails, never raw table writes. Cloudflare covers the edges: R2 for media, Workers for the web backend, Pages for the console.
The discipline shows in what I didn't build. Realtime is critical-but-narrow: clients subscribe only to what's actually on screen, and events ship as tiny envelopes rather than full payloads — no subscribe-to-everything, no five-second polling. Typing indicators and presence were deliberately deferred; they're nice, and nice isn't a launch reason. Both mobile apps are fully native — Compose on Android, SwiftUI on iOS — each respecting its platform instead of a lowest-common-denominator copy.
04 · Twenty days, with AI agents on the payroll
Here's the part I'm genuinely proud of: version 3 was rebuilt from scratch — all three surfaces, the backend, the governance model — in roughly twenty days, May 9 to May 28, 2026. Solo. The previous versions were thrown away entirely.
That speed didn't come from typing faster. It came from running AI coding agents the way I'd run an engineering team. Requirements came first: twenty-one structured interview rounds, consolidated into a single source of truth before serious code was written. Architecture got locked in nine ADRs. The work was sliced into missions, and every slice had to pass automated gates — SQL test suites, a sync-guard that rejects lazy polling patterns, an impact check tracing every change back to a requirement — before it could be committed. Fifty-eight SQL test suites and two hundred thirty UAT evidence documents later, the rule held: nothing lands unverified.
That's the actual skill on display. Anyone can ask an AI for code and get something plausible. Making agents produce production-grade, multi-platform software is a management problem — clear requirements, contracts, verification gates, and the judgment to know when output is wrong. It's the same discipline I apply to human teams, pointed at machines.
05 · What it shows
Tassenger is live: iOS 1.1.0 shipping on the App Store, Android in a 12-tester closed beta on Google Play with production access in sight, and the web console running on Cloudflare. Built in personal time, while running a national news platform as the day job.
For a hiring manager, the signal is layered: product judgment (two-mode design, trust modeled in the schema), real native mobile on both platforms, architecture chosen with a CFO's eye — and proof that I can direct AI tooling to serious, verified production outcomes, not demos. This is how I build when nobody is watching. It's also exactly how I'd lead an AI-augmented engineering team — because I already run one.