CrewTaskCrewTask - AI-Powered Family OS
CrewTask V2 (current version) - Full-stack AI Family OS. Built solo, part-time, in about 4 months (Feb - Jun 2026: ~3 months of development plus a month of Google Play release formalities) and iterating since, leveraging the foundational work of the first iteration.
Built by Marina Zorina · AI-native product builder · founder · Based in Playa del Carmen, Mexico
AI-native product builder. I take a vague idea to a shipped, reliable AI product and own it end to end - CrewTask, live on Google Play. Fifteen years in engineering - on why systems fail and how to prove they don't - is where the reliability comes from.
What I'm looking for
A founding or product-leadership role at an AI-first startup or venture studio - co-founder & CEO, head of product or product lead - where product strategy and building the thing yourself matter more than writing code by hand.
Roles I'm built for
Across my products the same thing repeats: I own what gets built and how we prove it works - strategy, architecture decisions, decomposition, acceptance criteria - and orchestrate AI to do the execution. That maps to several roles, each backed by something I actually shipped:
Co-founder & CEO who builds - take a product from a vague idea to a live app and own it end to end: direction, decisions, release. Proof: CrewTask, live on Google Play, built solo.
Head of Product / AI Product Lead - define LLM features through evals, structured outputs and explicit failure modes, and decide what is ready to ship. Proof: CrewTask's 49-example eval suite; Health Assistant, where every number is calculated by fixed formulas in code, and the LLM only explains the results in plain language.
Product Engineer, the product kind - the product person who does not stop at a spec: prototypes, evals, ships. Proof: CrewTask and Release Harness, both taken from idea to public release.
Entrepreneur in Residence - build a new venture from zero inside a studio. Proof: four products taken zero-to-one on my own; teams and systems I built keep working after I leave.
Currently full-remote from Mexico. Preferred, not required: GMT-5 ± 4 hours.
Client work
Stealth-mode HealthTech · Agentic AI · Silicon Valley · Aug 2026 - present. On the team of a company building agentic AI for healthcare; the company is in stealth, so it is not named here. I help shape how the product is released and work on the product itself: parts of the code, the testing framework, and a UI kit with a review of the product's screens. A close view of how an agentic AI product is built and shipped in a regulated industry.
The audit found green reports standing over runs whose checks never executed. I named the mechanism and closed it in CI, so a green report can no longer hide a red run. The layer I wired into every merge grew the daily report from 972 checks to 7,479. Every finding goes into a registry - 164 so far across product, AI behaviour, security, CI and process - each marked with how it was confirmed. Regulated industry: synthetic personas only, no real patient data, no production credentials.
Large fintech · Consultant · Apr 2022 - present. Consultant to a major bank.
An AI-powered Family OS for parents: say a chore as a voice note and it becomes structured, assignable tasks for the whole family. Built solo, part-time, in about four months.
- Three-tier Gemini orchestration (flash-lite, flash, pro) with a Groq fallback.
- An eval suite of 49 labeled examples in three languages, with an accuracy threshold checked before parser releases.
- 330 unit and 96 E2E tests, Sentry in production, offline-first sync.
Public mirror: github.com/sagittMZ/crewtask-public · the full story is below.
What CrewTask Is
CrewTask V2 is an AI-powered Family OS for parents who juggle work, kids, logistics, and endless mental load. You speak your chaos as a single voice note, and the system turns it into structured, assignable tasks for the whole family.
This is not just another todo app: it is a multi-tenant, mobile-ready system designed for real families with different devices and asymmetric responsibilities.
Where it is going: a community of families. Task trackers are everywhere. CrewTask is built around the family, not the individual, and the next layer is social: M-Tribe connects verified families nearby (KYC verification, opt-in location sharing), so a family can find trusted people around it - the support that relatives used to give when they lived close.
The fastest way to try it: scan with an Android phone, or open the store page.
Open in Google PlayMy Role
- Designed the product, architecture, and data model from scratch as a solo builder.
- Used LLMs (Cursor, Claude, Gemini) as collaborators, not magic: I orchestrated them, cross-validated outputs, and owned every decision.
- Implemented and verified RLS policies, offline-sync flows, and edge functions with failure-mode thinking.
- Shipped V2 in 4 months (part-time): ~3 months of development plus a month of Google Play release bureaucracy - while holding a full-time job and raising two kids.
Architecture at a Glance
Frontend & Mobile
- React 19 + TypeScript + Vite + TailwindCSS.
- Zustand for predictable client-side state management.
- Fully transitioned to Capacitor; the Android build is live in Google Play open testing, the iOS build is in preparation.
- Dual-path voice input: native OS STT as primary, with Gemini audio batch fallback for devices without Google services.
Backend & Data
- Supabase (PostgreSQL) with Row Level Security for multi-tenant families.
- Deno Edge Functions for AI task parsing, analytics, and business logic.
- Explicit schemas for users, families, tasks, members, and AI analytics.
- RLS policies designed with a security-first mindset, iteratively hardened using Supabase Advisor insights and verified against real-world failure modes.
AI & Orchestration
- Task parsing: native STT feeds voice-to-JSON via three-tier Gemini orchestration scaled by task complexity, with Groq as emergency fallback.
- Agentic orchestration and context scoping: development is driven by isolated AI agents with strict rulesets per surface.
- Human in the loop: manual cross-validation of AI architectural decisions across different contexts, leveraging over ten years of experience in product development teams.
AI Native Tooling & Orchestration
What orchestration means here. I do not ask one chatbot for code. I run a team of AI agents, each with its own scope, rules and context, and I run it the way a product owner runs a team: I set direction and priorities, decompose the work, define what "done" means, route each task to the right agent and model, and check the result against the goal before it ships. In industry terms: multi-agent orchestration, context engineering, agentic workflows, evals and guardrails, human-in-the-loop.
MCP connectors in practice. MCP (Model Context Protocol) gives agents direct, scoped access to real systems instead of me copying data between windows. Claude works against Supabase, Vercel and more on its own: it queries the live database, reads deployment logs and files tasks into CrewTask. Scheduled agents file tasks, send notifications and analyse the sources the work depends on, with no manual trigger - the same agent-integration pattern I am building into the product.
Built to work from anywhere. A phone-to-desktop bridge lets me drive the agents, review their output and approve releases away from the desk, so most of the work does not depend on where I am. The discipline behind it - commands, checks and session handoffs - is open source as Release Harness (see the tab above).
Implemented (MVP)
- AI Parser (Multi-tier Gemini 2.5 model orchestration for voice/text-to-task JSON).
- Smart Task Board (Mobile-first Kanban using Tailwind v4 + Zustand).
- Persistent quick-capture - an ongoing notification opens a voice composer; you speak and AI parses it, assigns it, and schedules it with priority.
- Push Notifications (Multi-channel: FCM for Android/iOS native, Web Push VAPID for browsers, Email via Resend).
- Offline-first sync UI (operability without a network with deferred synchronization).
- Automated Test Suite (Vitest: 330 unit tests in 35 files · Playwright: 96 E2E tests in 23 files · GitHub Actions CI on every PR).
- Error Monitoring (Sentry integration with breadcrumbs and source map support for production debugging).
- Google Play Open Testing - Android app is live in Google Play open testing; full public release is the next step.
6-12 Months Roadmap
- Lock-screen widget - quick task access; under exploration given Capacitor constraints, with the persistent notification shipping as the interim.
- Full Public Release on Google Play - now in open testing (anyone can join); full release next.
- Hands-free capture via OS assistants (exploring) - Siri Shortcuts / Intents (iOS) and Google Assistant App Actions (Android).
- iOS App - App Store submission and official release.
- Social Component (M-Tribe) - KYC (user verification) and location sharing for verified individuals.
- Two-way Google Calendar sync - temporarily disabled and being re-enabled.
Why This Project Matters
CrewTask started with my own family. We are two parents raising two kids abroad, with no grandparents or nannies nearby, and a big part of daily life happens in a language we are still learning. School, activities, groceries, bills, doctors - all of it lived in chats, voice notes and my head, and one phrase kept coming back: "you never told me". That phrase is where CrewTask started. The tools out there are built for one adult with an office schedule, not for a family. Many parents I talk to live the same way.
It began as a small helper for our home and grew into a real product, built the way production software is built:
- A real database with RLS, not a JSON file.
- A real mobile app, not just a desktop web page.
- Edge functions and AI parsing that can be audited and extended.
- Architecture that can grow into B2C or B2B2C without a rewrite.
Every product decision goes through one question: does this make family life clearer and calmer, or does it only look good in a screenshot?
Decision Stories (What I Can Walk You Through)
- Serverless Migration: I decided to build V2 on a Supabase + Vercel stack to completely offload server maintenance and database scaling to a BaaS. This allowed me, as a solo developer with strict deadlines, to focus exclusively on product business logic.
- Why Capacitor instead of native: One codebase, faster 0-to-1, enough native capabilities (push, widgets, voice triggers) via plugins without over-investing in Swift/Kotlin.
- How I use LLMs safely: Human in the loop, precise engineering prompts, and manual verification of the AI-generated architecture.
- When to rewrite: V1 had the core logic but accumulated UI/UX debt - responsiveness issues, design inconsistencies. The real shift was strategic: moving backend operations fully to Supabase BaaS to stop spending cycles on infra and focus entirely on product logic. Everything needed was already available out of the box.
- Supabase Auth Integration: Integration of Supabase Auth with Magic Links and seamless session recovery, drastically improving UX without unnecessary redirects.
How I Build
Across any stack. I orchestrate AI to ship in React, Python and Next.js - architecture, prompt pipelines, cross-model validation and the proof that it works are mine; the syntax is the model's job.