A full-stack cooking platform built solo, end to end — web, mobile, Postgres backend, and a server-side AI gateway. Shipped to a real domain, not a demo.
Nobody combines hands-free guided cooking, a real global recipe library, and AI that turns "here's a photo" or "here's my fridge" into something cookable tonight — plus a pantry that plans your week and writes the shopping list for you.
Cooksy does all four on one data model: web on Next.js 15, mobile in Flutter, Supabase Postgres underneath, and every AI call routed through a server-side gateway.
Full-screen, one step at a time, so the phone stays out of your hands while the pan is hot.
Snap a food photo and Claude identifies the dish, states its confidence, and returns a complete structured recipe with timed steps.
Give it what is in your kitchen plus constraints — time, diet, cuisine — and get three to five ranked ideas back.
"Plan my week with AI" builds a 7-day breakfast, lunch and dinner grid from your actual pantry, prioritizing ingredients about to expire.
Publish, save, rate and follow, on top of a curated global library rather than an empty feed waiting for users.
Solo across product decisions, architecture, data model, UI/UX, and both codebases — built AI-accelerated, as a repeatable method rather than a gimmick.
I run my own audits. A self-directed product and security review across the AI gateway, mobile, web, and Supabase advisors turned into a 17-task tiered fix plan — fully executed.
Design output gets the same treatment. Iteration runs on named feedback — "this reads as a generic SaaS table, make it editorial" — rather than accepting first-pass output.
Every AI call is a Next.js API route — never a client-side key. The pipeline below is what makes the AI features trustworthy and cheap to run, not just impressive.
Week planning fans out on top of it: one whole-week planning call plus seven parallel per-day recipe calls, assembled server-side.
Security-definer functions audited and fixed after a linter pass; foreign-key indexes added for performance.
Hourly missing-ingredient reminders, daily expiry checks, daily renewal notices — verified firing in production, not just scheduled.
The Flutter app mirrors the web event funnel event-for-event into a single PostHog dashboard.
Next.js on a VPS behind Docker, nginx and TLS via Cloudflare DNS-01 and certbot, custom domain, ~20s deploy through a small Python/paramiko script.
Free tier with AI quotas; Pro paid through eSewa and Khalti — Nepal's dominant mobile wallets — rather than a copy-paste Stripe integration.
A stale test migration had silently put every user on the paid plan with no expiry — both a revenue and an AI-cost leak.
Fix — Found and reverted the migration.
The seed script could never be safely run a second time, making environment rebuilds unreliable.
Fix — Made it idempotent so re-seeding is a non-event.
A security-definer view existed only in the live database, with nothing in version control that would recreate it.
Fix — Brought it back under migration so the live schema and the repo agree.
Next.js on a VPS behind Docker, nginx and TLS. Not a preview link.
Android build signed and FCM push confirmed landing on a physical device.
eSewa and Khalti, because that is who the market is.