🪴 Brain

Qamera AI

3 min read

Qamera AI

🗒️ Description

AI-powered virtual photo studio for e-commerce. Users select, compare, and approve AI-generated product photography and video — zero prompting required.

Built by 200IQ Labs PSA where I’m CTO & Co-Founder (30% equity).

🧩 Key features:

  • Packshot creation with automatic background removal
  • AI custom models (virtual mannequins)
  • AI scenery and backgrounds
  • Photo sessions (batches of 10 per config)
  • Video generation (HD, 5s clips)
  • Reel editing (auto-montages)
  • Style library (11+ pre-built visual styles)
  • Teams with role-based access and credits system

Target market

Primary niche: swimwear & lingerie brands (high sizing complexity, ad censorship requiring constant fresh variants). Secondary: broader fashion e-commerce.

Technology

Stack

  • Frontend/app: Next.js 16 (App Router) + React 19, Tailwind CSS 4, ui, scaffolded on MakerKit. Monorepo via Turborepo, pnpm.
  • Backend/data: Supabase (Postgres + Auth + Storage) with row-level security as the multi-tenant boundary. Airtable survives as a legacy bridge being migrated into Postgres.
  • Hosting: Vercel for the app; a self-hosted Docker stack on Hetzner runs the worker + queue. Docker for containers, Git (OpenSpec-style change proposals) for version control.
  • AI generation: multiple providers — BytePlus Seedream 4.0 (primary text/image-to-image, up to 4K), Google Gemini (image analysis + idea generation), Replicate (legacy fallback), Topaz Labs (upscaling).
  • Supporting services: Stripe (subscriptions + credit top-ups), Cloudflare Turnstile (CAPTCHA), Keystatic (blog/changelog CMS), Meta Ads Pixel + CAPI (conversion tracking), Sentry (errors), Vercel Analytics, ClickUp + email (dead-letter-queue alerts).

99% of code changes are made by AI coding agents (Agentic Coding). I design the environment, agents implement.

⚙️ Architecture know-how

  • Async job pipeline (HTTP variant of RabbitMQ). The web app publishes a run_due request to RabbitMQ; a thin Node worker on Hetzner consumes it, enforces platform + per-account provider limits, then calls back into the web app over authenticated HTTP (/api/internal/cg/process-one) to do the actual generation. Keeping the heavy logic in the web app avoids Vercel’s infinite-loop / timeout traps while the worker stays a stateless orchestrator.
  • Coalescing + watchdog. A 5-second bucket dedup (unique constraint on account_id, provider, bucket_ts) collapses bursts; a watchdog resets dispatched_at rows older than ~10 min to recover stuck jobs. State transitions use compare-and-swap for idempotent retries.
  • Credit system is application-first, ledger-based. Logic lives in a TypeScript service layer, not SQL functions: a credit_transactions history plus credit_reservations holds, guarded by advisory locks. Reservations are released on failure, confirmed on success — per-job pricing tiers (3–12 credits).
  • Repository + Service pattern, event-driven (in-memory now, designed to move onto the queue). RLS does authorization so loaders avoid manual auth checks; the admin client bypasses RLS only in rare, audited paths.

🎯 Go-to-market know-how

  • Platform-first pivot (2026-04): ~70% effort on e-commerce platform integrations (Saleor, Shoper, PrestaShop, then Shopify), ~30% founder-led direct sales in the swimwear/lingerie niche.
  • “Language of money” value selling: every artifact anchored to a concrete profit number + proof + before/after visual. A pre-discovery sales pipeline (account scoring → contact selection → opportunity FSM → opener → drafts → response monitoring → next-step) is itself implemented as a chain of agent skills over an Airtable CRM.

🔗 Links

Graph View

Backlinks